What is a GPU, and why does every company want them?

A GPU is a chip that can do thousands of small calculations at once. Modern AI is mostly that kind of arithmetic, so the companies building AI buy as many GPUs as they can get and race to put them to work.

If a company buys ten thousand GPUs, why can’t it use them the next day?

Because a GPU does nothing on its own. It needs a building, enough electricity, cooling, a fast network, software, testing and a team that knows it is there. Each of those comes from a different group with its own waiting list, and the slowest one sets the date.

Is this the same as being a data-center technician?

No. Technicians install, cable and repair hardware on site. Capacity and readiness operations is the coordinating role: it tracks every site and supplier, knows what is blocked and why, and drives the decisions that unblock it. Both are needed; one is hands-on, the other is mostly planning and communication.

What job titles should I search for?

Technical Program Manager with words like capacity, data center, infrastructure or deployment; Capacity Planner or Capacity Planning Manager; Hardware Deployment TPM; Infrastructure Program Manager; Commissioning Manager. The titles differ by company, so the pay section lists real examples with links.

Do these roles need a computer-science degree?

Usually not. Postings ask for experience running large, multi-team infrastructure programs, comfort with data and schedules, and the ability to make a decision when the information is incomplete. People come from construction, utilities, semiconductors, transit, aerospace and manufacturing as well as from cloud companies.

Are these roles remote?

Some are. Coordination roles that run forecasts, dependencies and readiness reviews across many sites are often remote or hybrid, because the work happens in dashboards and meetings. Commissioning and site-operations roles are on site. The remote section shows real postings of both kinds.

What does a strong candidate from outside AI bring?

The habits that make capacity productive are not AI-specific: readiness reviews that are decided before the date, dependencies tracked to an owner, acceptance defined before the test, and a single report leadership can act on. Someone who has carried a large physical program through commissioning has done this before, on different hardware.

What is the first thing a capacity and readiness lead should set up?

One honest picture of the funnel: how much capacity sits at each stage, per site, with the blocker and the owner for anything not moving. Most of the first wins come from making that picture visible, because the decisions it needs are usually cheap once someone can see them.

How do you measure success in this role?

Time from order to productive capacity, and the share of accepted capacity that is actually running work. Both should fall and rise respectively, batch after batch, and both should be visible to the people paying for the hardware.

Will this kind of work still exist in ten years?

The chips will change; the problem of turning scarce physical infrastructure into usable capacity will not. The outlook section gives the honest version, including the risks: spending cycles, overbuilding and hiring freezes.

← Writing

Productive Compute · Independent Research & Tools by Sina Nezhadian

AI Infrastructure Capacity & Readiness Operations: turning purchased GPUs into productive compute capacity

Companies are buying AI hardware faster than they can get it running. This page explains, in plain words, the work of closing that gap: what it is, why it matters, who needs it, what it pays, and what to watch.

GPU purchased ≠ productive GPU capacity

By Sina Nezhadian. Independent research: every outside claim links to its source and carries a date. Facts, inferences and open questions are labelled.

Published October 15, 2026 · Last updated October 15, 2026

01What is this?

When a company buys AI hardware, it buys boxes: graphics processing units (GPUs, the chips that do most AI arithmetic), the servers that hold them and the racks that hold the servers. None of it does any work until it has a building, enough electricity, cooling, a fast network, software, testing and a team that has been told it exists. AI Infrastructure Capacity & Readiness Operations is the work that connects the purchase to the result. It keeps one honest picture of how much capacity exists, how much is coming, what is blocking it, when it will be usable, and who has to decide what, to get it there.

Think of an airline that orders fifty planes. The planes are not capacity. Capacity is a plane with a crew, a gate, fuel, a maintenance sign-off and a route with passengers. Someone has to know which of the fifty are flying, which are parked, why, and what it would take to get the next one in the air. In AI infrastructure, that is this job.

02The pain: buying GPUs is not having capacity

Buying a GPU and having GPU capacity are different things, and the distance between them is measured in months. Hardware moves through a sequence of states before it does useful work. I call it the Productive Capacity Funnel. At every step, some of what was bought is stuck, waiting on something that is not a GPU: a transformer, a cooling loop, a network switch, a test, a signature, a decision.

The Productive Capacity Funnel. Only the last stage earns anything; everything above it is cost.

Where capacity gets stuck at each step

01
Ordered
The purchase order is signed. The hardware exists on a spreadsheet and in a supplier’s queue.
Where it sticks: Supplier lead times and allocation. The order is real; the delivery date is a forecast that can move.
An airline has signed for fifty planes. It has fifty delivery dates, not fifty aircraft.
02
Delivered
Crates arrive at the site, or at a warehouse near it.
Where it sticks: The building is not ready, so the crates wait. Delivery is often scheduled against the order date, not against the day the room can take them.
The planes land at an airport that has no gates finished yet, so they park on the apron.
03
Installed
Servers are in the racks, cabled to each other and to the network.
Where it sticks: Labour, rack space, cables and optics. A cluster needs thousands of parts that are not GPUs, and the cheap ones run out first.
The ovens are in the kitchen, bolted down, but the gas line is still being dug.
04
Infrastructure ready
Power, cooling and the network are live at the rack, at the capacity the hardware needs.
Where it sticks: Utility power, transformers, switchgear, cooling loops and network fabric. These have the longest lead times of anything in the funnel, and most of them are not controlled by the team that bought the GPUs.
A highway is paved and striped, but the on-ramps are not built, so no car can get on it.
05
Commissioned
Everything is powered on and tested together: every link, every GPU, every cooling loop, at full load.
Where it sticks: Failed burn-in tests, bad links and optics, firmware mismatches. Each failure means rework, and rework goes back through the queue.
The plane is fuelled, but a check finds a fault, and the fix needs a part that arrives next week.
06
Accepted
The owner formally accepts the capacity from the builder or supplier. Payment, warranty and support clocks start.
Where it sticks: Open punch-list items, missing test evidence and disagreements over what “done” means. Capacity can work perfectly and still be unaccepted.
The house is finished, but the inspector has not signed, so the family cannot move in.
07
Allocated
The capacity is assigned to a team, a customer or a workload.
Where it sticks: Quotas, scheduling and plain information: the teams who need it do not know it exists, or it is reserved for a project that has not started.
Fifty seats on the flight are empty because the booking system still shows the plane as grounded.
08
Productive
Real work is running on it, most of the time, at the performance that was paid for.
Where it sticks: Low utilisation, idle gaps between jobs, hardware failures and reserved-but-unused blocks. This is the only stage that earns anything.
The plane is flying full, on schedule. Everything before this was cost.

The hard part is that every step belongs to a different team: procurement, construction, facilities, networking, commissioning, finance, the platform team, the customer. Nobody owns the whole funnel unless someone is given the job of owning it.

03Why it costs so much

Idle hardware costs money twice. The money is already spent, or is being paid every month on a lease, and the work it was bought for is not happening. A simple way to see the size of it: multiply the number of GPUs by the days they sit idle and by the hours in a day. That is idle GPU-hours. Multiply by what an hour of GPU time is worth to you and you have a rough cost of the delay.

Idle-capacity calculator

Illustrative assumption: $3 per GPU-hour is a placeholder, not a market price. Replace it with the value of an hour on your hardware and contract.

Idle GPU-hours: 720,000
Estimated cost of the delay: $2,160,000
Per day of delay: $72,000

Idle GPU-hours = GPUs × days × 24. Estimated cost = idle GPU-hours × value per GPU-hour. Nothing you type here leaves your browser.

The default is deliberately round. What an hour is worth depends on the chip, the contract and what the work would have earned, so change it to whatever is true for you. The point is the shape of the number: a delay of weeks on a cluster of thousands is not a rounding error, it is the size of a serious project. That is why companies are hiring people whose whole job is to shorten it.

04Why demand is high right now

Three things are happening at once. The largest technology companies have said they will spend record sums on AI infrastructure. The electricity those buildings need is growing faster than the grid can add it. And the parts in between, from accelerators to transformers, have waiting lists. Each one alone would make readiness hard; together they make it the bottleneck. Every number below links to the document it came from.

FactOn Amazon’s July 30, 2026 second-quarter call, CEO Andy Jassy said the company now expects about $220 billion of cash capital expenditures in 2026, up from about $200 billion, because of the higher cost of memory, and that it will still not have enough capacity to meet all demand in 2026 or, he expects, in 2027. (Amazon, via Investing.com, 2026-07-30)In plain words: Amazon raised its yearly building budget by $20 billion just because memory chips got more expensive, and it still expects to run short of capacity through next year.
FactOn its July 22, 2026 second-quarter call, Alphabet raised its full-year 2026 capital-spending guidance to $195 billion to $205 billion, from $180 billion to $190 billion, with CFO Anat Ashkenazi attributing the increase mainly to faster delivery of capacity in a supply-constrained environment. (Alphabet, via Investing.com, 2026-07-22)In plain words: Google’s parent company plans to spend about $200 billion in one year on computers and data centers, and raised the number mainly because capacity is being delivered faster, while supply is still short.
FactThe International Energy Agency said on April 16, 2026 that electricity demand from data centres grew 17% in 2025 and is set to double by 2030, and that AI deployment is increasingly running into physical bottlenecks: tighter supply of gas turbines and transformers, advanced chips, and slow grid connections. (International Energy Agency, 2026-04-16)In plain words: The world’s energy watchdog says data centers will need twice as much electricity by 2030, and the things needed to deliver it are already in short supply.
FactLawrence Berkeley National Laboratory’s Queued Up 2026 report says about 2,061 gigawatts of generation and storage were waiting for U.S. grid interconnection at the end of 2025, and that the median project built in 2025 took 61 months from interconnection request to commercial operation, against 36 months in 2015. (Lawrence Berkeley National Laboratory (U.S. Department of Energy), 2026-06)In plain words: A new U.S. power plant now typically takes about five years from asking to connect to the grid to actually producing power, so the electricity a data center needs can take far longer to arrive than the computers do.
FactOn TSMC’s October 15, 2026 third-quarter call, CEO C.C. Wei said the gap between demand and supply for its chips is still quite big, acknowledged a small shortage in back-end capacity, and the company reiterated a 2026 capital budget of $60 billion to $64 billion. (TSMC, via Investing.com, 2026-10-15)In plain words: TSMC, the chip manufacturer, says demand for its chips is still far above what it can supply, even with a 2026 capital budget of $60 billion to $64 billion.
FactIn its September 30, 2026 prepared remarks, Micron said it has completed agreements for the vast majority of its calendar 2027 high-bandwidth memory supply, with significant price increases, and expects memory supply to be much tighter in 2027 and 2028 than in 2026. (Micron Technology, 2026-09-30)In plain words: The special memory chips that sit next to every AI GPU are mostly spoken for through next year, at higher prices, so a GPU order can be held up by its memory, not just the GPU.
FactGE Vernova reported on July 22, 2026 that its gas-power equipment backlog plus slot reservations grew from 100 to 116 gigawatts in the second quarter, that it expects at least 125 gigawatts under contract by year-end, and that it is on track for 20 gigawatts a year of gas-turbine output, with actions under way to produce 30 gigawatts in 2030. (GE Vernova via SEC EDGAR, 2026-07-22)In plain words: A major turbine maker has well over 100 gigawatts of orders and reservations against a factory output of about 20 gigawatts a year, so a data center that wants its own power plant must book a place in line years ahead.

05The 10 companies and industries that need this most

Ten companies sit at the centre of the buildout. For each one, the same six short blocks: why the problem matters to them, what they are publicly doing about it, how they do it, the latest dated updates, who leads it where a public source names someone, and the real job titles that do the work, with pay ranges only where the company’s own posting prints one.

A rule for reading this section: a Fact is what the company or a cited source publicly said. An Inference is what it likely means. A Question is what to watch. Nothing here accuses any company of waste or failure, and no internal number is invented; if it is not public, it is not here.

NVIDIA

The company that sells the chips says supply, not demand, limits its growth, and now talks about land, power and buildings as much as silicon.

Why this problem matters to them

NVIDIA sells the hardware, and its own numbers show that the hard part is turning that hardware into working capacity. In the quarter ended July 26, 2026 its data-center revenue was $89.0 billion, and the company said its compute is fully used across every cloud it serves. Its CFO said supply will stay the bottleneck at least through the end of fiscal 2028. So every rack that waits in a warehouse, or sits in a building with no power or network yet, is output nobody can replace. That is why NVIDIA now talks about land, power, building shells, cooling, networking and deployment teams, not only chips. (NVIDIA, 2026-08-26) (NVIDIA, 2026-08-26)

It is like an airline with more passengers than seats. Buying more planes does not help on its own if the airports have no gates, fuel or crews ready on the day the planes arrive.

Show what NVIDIA is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactShipping the new Vera Rubin generation: production shipments began in August 2026, racks are already running at five cloud partners, and NVIDIA expects it to be about a fifth of data-center revenue in the quarter ending October 2026. (NVIDIA, 2026-08-26) (NVIDIA, 2026-08-26)
FactHelping partners lock in land, power and buildings before the hardware arrives: a SoftBank Energy campus at Portsmouth whose first phase is expected to support 4.25 gigawatts for OpenAI, and an Australian program with eight partners targeting up to 2 gigawatts by 2027. (NVIDIA, 2026-08-26) (NVIDIA, 2026-09-09)
FactGrowing neocloud capacity with DSX reference designs: NVIDIA says these partners should end 2026 with 8 gigawatts installed, up from about 3 gigawatts a year earlier. (NVIDIA, 2026-08-26) (NVIDIA, 2026-03-16)
FactOpening its rack designs and pushing power and network standards so others can build faster: Vera Rubin rack designs given to the Open Compute Project, plans for an 800-volt DC AI-factory platform, Meta adopting its newest switches, and a new optics format with Cisco. (NVIDIA, 2026-10-12) (NVIDIA, 2026-10-12)
FactHiring for the last mile itself: a capacity-and-deployment program manager, a DGX Cloud capacity-management engineer, a first-time-deployment manager, a Texas-remote field build engineer, and customer program managers who run AI-factory build-outs with cloud providers. (NVIDIA careers, 2026-10-14) (NVIDIA careers, 2026-10-08) (NVIDIA careers, 2026-09-29) (NVIDIA careers, 2026-10-14) (NVIDIA careers, 2026-07-20)

How they do it

FactMostly through other people’s data centers: hyperscalers and neoclouds run the racks, and AWS alone is adding 2 million more NVIDIA GPUs between the quarter ending October 2026 and the quarter ending July 2028. (NVIDIA, 2026-08-26) (NVIDIA, 2026-08-26)
FactReference designs instead of custom builds: the DSX AI-factory reference design gives partners a repeatable plan for power, cooling and racks, and the October 2026 Open Compute contributions extend it to open rack designs. (NVIDIA, 2026-03-16) (NVIDIA, 2026-10-12)
FactBuying supply before demand lands: inventory rose to $32 billion ahead of the Vera Rubin launch, and the CFO called multi-year supply commitments essential to the ramp. (NVIDIA, 2026-08-26)
FactPower and site partnerships rather than owning power plants: land, power and shell through SoftBank Energy at Portsmouth, Australian partners expanding capacity for DSX factories, and an 800-volt DC power design for future sites. (NVIDIA, 2026-08-26) (NVIDIA, 2026-09-09) (NVIDIA, 2026-10-12)
FactMeasuring an AI factory by what it produces: NVIDIA’s own glossary says the product is intelligence measured in tokens, and tokens only exist while the hardware is running. The company also says each gigawatt is worth more to it with every generation. (NVIDIA, undated, read 2026-10-15) (NVIDIA, 2026-08-26)

Latest updates, as of October 2026

2026-10-14Posted a Technical Program Manager, Infrastructure Capacity and Deployment role in Santa Clara and a Texas-remote field engineer for AI-infrastructure build execution, both with printed pay ranges. (NVIDIA careers, 2026-10-14)
2026-10-12At the OCP Global Summit, contributed Vera Rubin rack and tray designs to the Open Compute Project; partners unveiled plans for an 800-volt DC AI-factory platform. (NVIDIA, 2026-10-12)
2026-10-12With Cisco, announced the USFP optics form factor so AI factories can add compute without adding network tiers. (NVIDIA, 2026-10-12)
2026-09-09Announced a program with eight Australian partners to expand land, power and shell capacity for DSX AI factories, up to 2 gigawatts by 2027. (NVIDIA, 2026-09-09)
2026-08-26Second-quarter results: $96.2 billion revenue, $89.0 billion data center; the CFO said Vera Rubin shipments began in August and that supply will stay a bottleneck at least through fiscal 2028. (NVIDIA, 2026-08-26)
2026-08-26Said AWS will deploy an additional 2 million NVIDIA GPUs through the quarter ending July 2028. (NVIDIA, 2026-08-26)

Who leads it

Jensen Huang, Founder, President and CEO — sets the AI-factory strategy and speaks for supply and demand on earnings calls (NVIDIA, undated, read 2026-10-15)
Debora Shoquist, EVP, Operations — NVIDIA’s operations organization, as listed on the management page (NVIDIA, undated, read 2026-10-15)
Ian Buck, Vice President of Hyperscale and High-Performance Computing — the hyperscale business; named in NVIDIA’s post on opening the Vera Rubin rack designs (NVIDIA, 2026-10-12)
Raj Mirpuri, Vice President of Global AI Clouds and Infrastructure Ecosystem — the cloud-partner and data-center ecosystem; quoted in the Australia capacity release (NVIDIA, 2026-09-09)

The roles that do this work there

Technical Program Manager, Infrastructure Capacity and Deployment
Leads the full lifecycle of infrastructure demand: intake and prioritization, capacity forecasting, execution reporting and operational health across data centers, offices and other infrastructure.
Santa Clara, CA · remote policy not stated · 168,000 USD – 258,750 USD for Level 4, and 200,000 USD – 322,000 USD for Level 5 (base salary per year, plus equity and benefits) · posted 2026-10-14
Senior Customer Technical Program Manager – AI Datacenter
Runs deployment programs with cloud providers and their manufacturing partners for GB300 NVL72 systems and clears blocking issues in AI-factory build-outs.
Santa Clara, CA (additional location: US, Remote) · US-remote listed as an additional location · 168,000 USD – 258,750 USD for Level 4, and 200,000 USD – 322,000 USD for Level 5 (base salary per year, plus equity and benefits) · posted 2026-07-20
Manager, First Time Deployment – NVIS
Leads the team that does the first deployment of new NVIDIA systems: site preparation, install and bring-up evidence, validation data and feedback to design engineering.
Santa Clara, CA · the posting says it is open to a remote work location · 184,000 USD – 287,500 USD for Level 3, and 216,000 USD – 345,000 USD for Level 4 (base salary per year, plus equity and benefits) · posted 2026-09-29
Senior Field Applications Engineer, AI Infrastructure Build – Execution
Gives hands-on support at customer and partner sites during AI-factory builds and turns field findings into repeatable build guidelines.
Texas, remote (additional location: Houston, TX) · listed as US, TX, Remote; on site during build, bring-up and validation windows · 132,000 USD – 207,000 USD for Level 3, and 160,000 USD – 253,000 USD for Level 4 (base salary per year, plus equity and benefits) · posted 2026-10-14
Senior Customer Technical Program Manager – Data Center Quality and Sustaining
Owns quality and sustaining programs with manufacturing partners across large cloud deployments, including high-profile customer issues and a joint quality-assurance strategy.
Santa Clara, CA (additional location: Austin, TX) · remote policy not stated · 200,000 USD – 322,000 USD (base salary per year, plus equity and benefits) · posted 2026-10-14
Senior Technical Program Manager, Data Center – Engineering Operations
Drives cross-functional engineering-operations programs for the data-center server platforms used by cloud providers.
Santa Clara, CA (additional location: Hillsboro, OR) · remote policy not stated · 168,000 USD – 258,750 USD for Level 4, and 200,000 USD – 322,000 USD for Level 5 (base salary per year, plus equity and benefits) · posted 2026-10-11
Director, Boards and Systems Supply Planning
Directs supply planning for board and system products, matching what can be built against what customers have ordered.
Santa Clara, CA · remote policy not stated · 216,000 USD – 339,250 USD (base salary per year, plus equity and benefits) · posted 2026-10-12

What it likely means, and what to watch

InferenceNVIDIA’s own supply-constrained outlook suggests the binding constraint has moved beyond the chip to the whole chain, including memory, racks and the land, power and shells partners must have ready, which is likely why its releases now talk about land, power and shell.
InferenceThe expected jump from about 3 to 8 gigawatts of neocloud capacity in one year suggests NVIDIA is leaning on neoclouds and reference designs to turn shipped systems into usable capacity faster than traditional build cycles allow.
InferenceThe cluster of new deployment and capacity postings suggests NVIDIA is building an in-house function for the last mile of deployment and utilization rather than leaving it entirely to customers.

QuestionWill the next quarterly report, due after the quarter ending in late October 2026, keep the “supply bottleneck through fiscal 2028” language?
QuestionWill neocloud partners actually reach 8 gigawatts installed by the end of 2026, and how much of that will be energized and in use rather than shipped?
QuestionWhich link does NVIDIA name as the bottleneck next: memory pricing, packaging, power or building shells?

OpenAI

A lab whose finance chief says revenue tracks the compute it can use, with far more capacity contracted than live.

Why this problem matters to them

OpenAI’s own finance chief says the company’s revenue has tracked the compute it could actually use, and that compute is its scarcest resource. OpenAI ended 2025 with about 1.9 gigawatts of compute in use; by April 2026 it said it had lined up more than 10 gigawatts of U.S. capacity. The gap between those two numbers is the work this page is about: turning signed contracts and delivered racks into power, cooling, networks and tested clusters that models can run on. (OpenAI, 2026-01-18) (OpenAI, 2026-04-29)

It is like an airline that has bought a fleet of jets but is still waiting on runways, fuel lines and trained crews. The planes are real and paid for, but they earn money only once all of that is ready.

Show what OpenAI is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactAnnounced in April 2026 that it had passed its 10-gigawatt U.S. goal years early, adding more than 3 gigawatts in 90 days; Data Center Dynamics judged this likely to be contracted rather than live capacity, while the last reported figure for compute in use was about 1.9 gigawatts at the end of 2025. (OpenAI, 2026-04-29) (Data Center Dynamics, 2026-05-01) (OpenAI, 2026-01-18)
FactBuilding Stargate sites with Oracle and SoftBank across Texas, New Mexico, Wisconsin, Ohio and Michigan; the flagship Abilene site is live and trained GPT-5.5. (OpenAI, 2025-09-23) (OpenAI, 2025-10-30) (OpenAI, 2026-04-29)
FactStanding up an in-house delivery and commissioning organization for self-built campuses, and hiring in August–October 2026 for commissioning program managers, a hardware supply-chain program manager, a first hardware quality and reliability engineer, and a model-deployment-and-capacity program manager. (OpenAI careers, 2026-08-10) (OpenAI careers, 2026-08-10) (OpenAI careers, 2026-09-24) (OpenAI careers, 2026-09-28) (OpenAI careers, 2026-09-02)
FactWidening where capacity comes from: about 2 gigawatts of AWS Trainium agreed in February 2026, Malaysian capacity from Firmus in September 2026, and talks for smaller 20-to-30-megawatt deployments in the U.S. and the Nordics. (Amazon via SEC EDGAR, 2026-02-27) (Data Center Dynamics, 2026-09-08) (Data Center Dynamics, 2026-09-23)

How they do it

FactCloud partners carry most of the load: Oracle for Stargate, Microsoft (an incremental $250 billion of Azure services) and AWS (about 2 gigawatts of Trainium, with Amazon investing $50 billion in OpenAI). (OpenAI, 2025-09-23) (OpenAI, 2025-10-28) (Amazon via SEC EDGAR, 2026-02-27)
FactSeveral chip suppliers with dated delivery milestones: at least 10 gigawatts of NVIDIA systems from the second half of 2026, 6 gigawatts of AMD GPUs from the second half of 2026, and 750 megawatts of Cerebras systems from 2026. (OpenAI, 2025-09-22) (OpenAI, 2025-10-06) (Cerebras, 2026-01-14)
FactCustom silicon: OpenAI designs its own accelerators and racks with Broadcom, with 10 gigawatts targeted to start in the second half of 2026 and finish by the end of 2029. (OpenAI, 2025-10-13)
FactSelf-built campuses delivered with developer partners, with OpenAI providing commissioning leadership and governance while commissioning partners provide field and test engineering. (OpenAI, 2025-10-30) (OpenAI careers, 2026-08-10) (OpenAI careers, 2026-08-10)
FactPower arranged site by site, for example DTE Energy serving the Michigan campus from existing excess transmission capacity. (OpenAI, 2025-10-30) (OpenAI, 2026-04-29)

Latest updates, as of October 2026

2026-09-28Posted a first-hire Data Center Hardware Quality & Reliability Engineer to own the loop from field failures to verified root cause, spares forecasts and upstream fixes. (OpenAI careers, 2026-09-28)
2026-09-24Posted a Supply Chain Operations Program Manager for new hardware introductions and production ramps across its AI infrastructure systems. (OpenAI careers, 2026-09-24)
2026-09-23Reported to be in talks for 20-to-30-megawatt capacity deals in the U.S. and the Nordics, likely for inference, and said it is building a diversified compute portfolio. (Data Center Dynamics, 2026-09-23)
2026-09-08Firmus will supply OpenAI with compute from two data centers in Malaysia, with OpenAI as anchor customer. (Data Center Dynamics, 2026-09-08)
2026-09-02Posted Technical Program Manager, Model Deployment & Capacity, connecting demand forecasting and capacity allocation with model readiness and rollout. (OpenAI careers, 2026-09-02)

Who leads it

Sarah Friar, Chief Financial Officer — wrote the January 2026 post on compute as the scarcest resource and revenue tracking available compute (OpenAI, 2026-01-18)
Greg Brockman, Co-founder and President — quoted on OpenAI’s own announcements of the NVIDIA, AMD and Broadcom deployments (OpenAI, 2025-09-22)
Commissioning organization, Industrial Compute — makes sure self-built and partner-delivered infrastructure is tested, validated, integrated and handed into reliable operations (OpenAI careers, 2026-08-10)

The roles that do this work there

Technical Program Manager, Model Deployment & Capacity
Runs the operating system for ChatGPT capacity and model deployment, connecting demand forecasting and capacity allocation with model readiness and rollout.
San Francisco · hybrid, three days a week in the office · $207K – $445K, plus equity (salary as printed on the board) · posted 2026-09-02
Technical Program Manager, Compute Infrastructure
Owns end-to-end delivery of large GPU clusters across external providers, running hardware, networking, power and cooling work in parallel to bring production-ready capacity online.
San Francisco · hybrid, three days a week in the office · $257K – $335K, plus equity (salary as printed on the board) · posted 2026-03-12
Supply Chain Operations Program Manager, AI Infrastructure
Single-threaded operational leader for new hardware introductions and production ramps, holding contract manufacturers and integrators to the critical path.
San Francisco · hybrid; supports factory builds and field deployments with travel · $177K – $285K, plus equity (salary as printed on the board) · posted 2026-09-24
Commissioning Program Manager, Infrastructure Delivery
Builds the common standards, tools and reporting that let commissioning run consistently across self-build campuses.
San Francisco or a self-build campus · not a remote position, as the posting states · $177K – $285K, plus equity (salary as printed on the board) · posted 2026-08-10
Commissioning Project Lead, Data Center Delivery
Owns commissioning strategy and execution for a large self-build data center, from construction planning through integrated systems testing to final turnover.
San Francisco; Seattle · on site; hybrid during planning, then full time on site through startup, testing and turnover · $226K – $285K, plus equity (salary as printed on the board) · posted 2026-08-10
Data Center Infrastructure Engineering Program Manager (Network and Whitespace)
Turns data-center strategy into executable programs for network integration and data-hall design and deployment, across construction, commissioning and operations.
San Francisco; Seattle; US – Remote · hybrid, with US-remote listed; works across office, lab, construction and operating sites · $226K – $285K, plus equity (salary as printed on the board) · posted 2026-04-20
Data Center Hardware Quality & Reliability Engineer
First hire to own the hardware quality and reliability loop across third-party and first-party infrastructure: field failures to containment, verified root cause, spares forecasts and upstream fixes.
San Francisco · hybrid · $226K – $285K, plus equity (salary as printed on the board) · posted 2026-09-28

What it likely means, and what to watch

InferenceThe distance between roughly 1.9 gigawatts in use at the end of 2025 and more than 10 gigawatts contracted by April 2026 suggests OpenAI’s next two years are mostly a readiness problem rather than a purchasing problem.
InferenceThe August–September 2026 postings for commissioning program managers, a supply-chain program manager and a first hardware quality and reliability engineer suggest OpenAI is building its own delivery and field-quality discipline rather than leaving it entirely to partners.
InferenceSmaller 20-to-30-megawatt deals and Malaysian capacity suggest OpenAI is placing inference capacity closer to users at sites that can be ready sooner than gigawatt campuses.

QuestionHow much of the 10+ gigawatts contracted will be energized, accepted and running workloads by the end of 2026, and will OpenAI report live capacity separately from contracted capacity?
QuestionWill the first NVIDIA Vera Rubin, AMD MI450 and Broadcom custom racks arrive on the second-half-2026 schedules the companies published, and how long from rack delivery to productive use?
QuestionHow will the mix between cloud, neocloud and self-build change, and what are the timelines of the Wisconsin, Michigan, Ohio and New Mexico sites?

Anthropic

An AI lab whose chips come from five suppliers and whose buildings come from partners, now hiring a capacity-delivery team of its own.

Why this problem matters to them

Anthropic sells access to Claude, so an hour a chip sits idle is an hour of product it cannot sell. In May 2026 its CEO said the company had planned for 10-fold growth, saw 80-fold growth on an annualized basis in the first quarter, and had difficulties with compute as a result. It has signed for many gigawatts of chips from Amazon, Google, NVIDIA, AMD and SpaceX, but much of it is scheduled for 2027 or later, and only about one gigawatt of the Amazon capacity is due by the end of 2026. The hard part is in between: getting buildings powered, racks installed, networks turned up and chips tested so they can serve customers. That is why its own postings describe a job that takes a site from construction kickoff to ready for service. (CNBC, 2026-05-06) (Anthropic, 2026-04-20) (Anthropic careers, 2026-09-22)

It is like an airline that has ordered hundreds of planes from several makers. The planes earn money only once runways, gates, pilots and maintenance crews are ready, and the order book says nothing about when that happens.

Show what Anthropic is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactLocking in multi-gigawatt chip supply from several vendors at once: up to 5 gigawatts from Amazon, multiple gigawatts of Google TPUs from 2027, up to 2 gigawatts of AMD GPUs from the first half of 2027, up to a gigawatt on Azure, and all of SpaceX’s Colossus 1 from May 2026. (Anthropic, 2026-04-20) (Anthropic, 2026-04-06) (AMD, 2026-07-22) (Microsoft, 2025-11-18) (Anthropic, 2026-05-06)
FactShifting from renting cloud capacity toward long leases of purpose-built sites: $50 billion of custom data centers with Fluidstack in Texas and New York, a 20-year lease of a roughly 400-megawatt TeraWulf site in Kentucky, and a first Australian site in Queensland. (Fluidstack, 2025-11-12) (CNBC, 2026-07-06) (ABC News (Australia), 2026-09-16)
FactBuilding an in-house Data Center Operations and Capacity Engineering organization: open roles cover capacity deployment leads, partner-site operations leads, demand planners, delivery reporting and controls, compute-supply reliability engineers and data-center program managers. (Anthropic careers, 2026-09-22) (Anthropic careers, 2026-08-21) (Anthropic careers, 2026-08-11) (Anthropic careers, 2026-08-24) (Anthropic careers, 2026-10-05) (Anthropic careers, 2026-10-07)
FactGoing international: Compute Country Leads for Japan, Korea and Canada are each accountable for bringing gigawatts online. (Anthropic careers, 2026-08-11)
FactTreating local communities as part of delivery: on October 15, 2026 it posted Community Engagement Manager roles for data centers in Louisiana, New York, and Virginia and West Virginia, joining earlier Texas, Canada and Australia postings from August 19, 2026. (Anthropic careers, 2026-10-15)

How they do it

FactPartners’ custom chips rather than its own silicon: over one million Amazon Trainium2 chips in use, Google TPUs under an agreement with Google and Broadcom, NVIDIA systems on Microsoft Azure, and AMD Instinct GPUs from 2027. (Anthropic, 2026-04-20) (Anthropic, 2026-04-06) (Microsoft, 2025-11-18) (AMD, 2026-07-22)
FactNeocloud and direct-lease deals for buildings and power: Fluidstack builds custom data centers for it, TeraWulf and Nscale lease sites to it, and SpaceX supplies a whole existing data center. Anthropic is the tenant and the workload owner, not the builder. (Fluidstack, 2025-11-12) (CNBC, 2026-07-06) (Anthropic, 2026-05-06)
FactA partner-operated site model: Anthropic staff set the standards and verify the work while a site operations partner does the day-to-day data-hall work, and deployment leads are on site for power-on, network turn-up, burn-in and acceptance. (Anthropic careers, 2026-08-21) (Anthropic careers, 2026-09-22)
FactCapacity Engineering owns the single capacity data layer and the planning tools, and reliability engineers build the power and availability models that decide which sites to lease and how far to oversubscribe. (Anthropic careers, 2026-08-11) (Anthropic careers, 2026-10-05) (Anthropic careers, 2026-09-11)
FactProgram management from construction through commissioning, with repeatable playbooks and one milestone view of every project from deal close to operational handover. (Anthropic careers, 2026-10-07) (Anthropic careers, 2026-08-24) (Anthropic careers, 2026-03-02)

Latest updates, as of October 2026

2026-10-15New postings for Community Engagement Manager, Data Centers in Louisiana, New York, and Virginia and West Virginia, which signal the states where Anthropic expects to site, build and operate data centers. (Anthropic careers, 2026-10-15)
2026-10-05Posted a Data Center Engineer, Compute Supply role whose power and availability models will shape which sites it leases and how far it oversubscribes. (Anthropic careers, 2026-10-05)
2026-09-22Posted a Capacity Deployment Lead role that runs new capacity from construction kickoff to ready for service across partner-operated sites, with 50%+ travel. (Anthropic careers, 2026-09-22)
2026-09-16ABC News reports Anthropic signed to use part of a proposed $32 billion data centre near Dalby, Queensland, aiming to start in 2027, subject to approvals. (ABC News (Australia), 2026-09-16)

Who leads it

Dario Amodei, CEO and co-founder — leads the company and publicly fronts its compute agreements, including the Amazon expansion (Anthropic, 2026-04-20)
Krishna Rao, Chief Financial Officer — finance; spoke for the company on the Google and Broadcom TPU expansion (Anthropic, 2026-04-06)
Data Center Operations (DCO) and Capacity Engineering teams — bringing new capacity online across partner-operated sites and keeping the fleet accounted for and well-utilized; no public source names the individuals who lead them (Anthropic careers, 2026-09-22)

The roles that do this work there

Capacity Deployment Lead – Data Center Operations
Brings new compute capacity online at partner-operated data centers, on site from infrastructure delivery through power-on, network turn-up, burn-in, validation and acceptance.
Remote-Friendly, United States · Remote-Friendly; 50%+ travel and extended on-site work during delivery and turn-up windows · $320,000—$405,000 USD (annual salary) · posted 2026-09-22
Infrastructure Capacity Planner, Demand Planning
Plans compute demand across a fleet of clouds, neoclouds and on-prem sites so supply deals, delivery and utilization line up.
San Francisco, CA | New York City, NY · hybrid; in an office at least 25% of the time · $320,000—$405,000 USD (annual salary) · posted 2026-08-11
Technical Program Manager, Data Center Infrastructure
Drives data-center programs from construction through commissioning and operations, building repeatable playbooks across concurrent site deployments.
San Francisco, CA | New York City, NY | Seattle, WA · hybrid; at least 25% office time, with 20% travel to sites · $365,000—$435,000 USD (annual salary) · posted 2025-11-26, updated 2026-10-07
Data Center Operations Lead – Partner Site Operations
Owns deployment velocity, availability and incident response at partner-operated sites by directing and assessing the vendor’s on-site teams.
Remote-Friendly (travel required) | San Francisco, CA · Remote-Friendly with travel; the standard hybrid line is also printed · $320,000—$405,000 USD (annual salary) · posted 2026-08-21
Reporting and Controls Lead, Data Center Capacity Delivery
Owns the single view of every project in the delivery portfolio, from deal close through floor access and operational handover.
Remote-Friendly, United States · Remote-Friendly; the standard hybrid line is also printed · $320,000—$405,000 USD (annual salary) · posted 2026-08-24
Technical Program Manager, Compute
Runs programs across the compute lifecycle, from how supply is procured and brought online to how capacity is allocated and used.
San Francisco, CA | New York City, NY | Seattle, WA · hybrid; in an office at least 25% of the time · $290,000—$365,000 USD (annual salary) · posted 2026-03-02
Senior Engineering Manager, Capacity Engineering
Leads the team that keeps all infrastructure resources accounted for, well-utilized and efficiently allocated across first-party and third-party compute.
San Francisco, CA | New York City, NY | Seattle, WA · on site five days a week, as the posting states · $405,000—$485,000 USD (annual salary) · posted 2026-09-11

What it likely means, and what to watch

InferenceAnthropic’s postings describe a fleet that spans clouds, neoclouds and on-prem sites across several chip families, so its readiness problem is multi-partner and multi-platform rather than one cloud contract.
InferenceMuch of the announced capacity is scheduled for 2027 or later, so the capacity that serves customers in 2026 depends mainly on Amazon, the earlier Google TPU deal, SpaceX’s Colossus 1 and Azure.
InferencePhrases such as “construction kickoff to ready for service” and 50%+ travel suggest Anthropic treats on-site acceptance, not a dashboard, as the proof that purchased capacity is productive.

QuestionHow much of the nearly one gigawatt of Trainium capacity promised for the end of 2026 is live and serving Claude by December?
QuestionAre the Fluidstack sites in Texas and New York, due online throughout 2026, in service by year-end? No public status update was found.
QuestionWill Anthropic ever disclose fleet utilization or ready-for-service metrics, given that its Capacity Engineering team says it owns the canonical capacity data?

Google (Alphabet / Google Cloud)

The one hyperscaler with its own chip at scale, now renting outside GPU capacity as a bridge while it finishes its own buildings.

Why this problem matters to them

Alphabet now expects to spend $195 billion to $205 billion on capital expenditures in 2026, and it spent $44.9 billion in the second quarter alone. Most of that money buys chips, servers and buildings, which earn revenue only once the hardware is installed, powered, cooled, connected and handed to a paying workload. In July 2026 Google’s finance chief said supply constraints meant the company would rent outside capacity as a bridge while it finishes building its own. The gap between money spent and capacity that is actually live is the problem this page is about. (Alphabet, via Investing.com, 2026-07-22) (Alphabet via SEC EDGAR, 2026-07-22) (CNBC, 2026-07-22)

It is like a city that buys a thousand new buses while the garage, the fuel depot and the drivers are still a year away. Until those arrive, the buses sit parked and carry no passengers.

Show what Google (Alphabet / Google Cloud) is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactSpending more and faster: Alphabet raised its 2026 capital-spending range to $195 billion to $205 billion in July 2026, saying the increase was mainly about delivering capacity sooner; it spent $80.6 billion in the first half of 2026 and reported $513.9 billion of Google Cloud backlog. (Alphabet, via Investing.com, 2026-07-22) (Alphabet via SEC EDGAR, 2026-07-22) (Alphabet via SEC EDGAR, 2026-07)
FactRenting outside capacity as a bridge: the CFO said Google would expand third-party capacity in the third quarter while internal builds catch up, and a SpaceX filing shows Google agreed to pay $920 million a month from October 2026 for about 110,000 NVIDIA GPUs, with a September 30, 2026 delivery deadline. (CNBC, 2026-07-22) (SpaceX via SEC EDGAR, 2026-06-05)
FactTurning its own chips into a product: Ironwood became generally available on March 31, 2026, eighth-generation TPUs were described in April, Anthropic signed for multiple gigawatts of TPU capacity starting in 2027, and in September Kurian said the TPU business is more than twice the next hyperscaler’s accelerator business. (Google Cloud documentation, 2026-03-31) (Google Cloud, 2026-04-22) (Google Cloud, 2026-04-06) (Investing.com, 2026-09-08)
FactLocking in power years ahead: in October 2026 alone, 890 megawatts of new nuclear capacity with Constellation (first uprate by 2028), a 702-megawatt fuel-cell plant for Google in Wyoming (full operation by the end of 2028) and an approved 390-megawatt utility plan in Indiana; in Finland, a 22-year nuclear life-extension deal and a 94-megawatt battery accompany the 13 billion euro build. (Google Cloud, 2026-10-06) (ESG Today, 2026-10-15) (Indiana Public Media, 2026-10-12) (Google, 2026-09-09)
FactHiring people whose job is readiness, not purchasing: a Capacity Delivery program manager covering site selection through commissioning and handover, a leader who owns the Alphabet-wide TPU and GPU plan of record, and capacity planners whose team exists to unlock stranded power and floor space. (Google careers, undated, read 2026-10-15) (Google careers, undated, read 2026-10-15) (Google careers, undated, read 2026-10-15)

How they do it

FactOwn data centers plus third-party colocation: the Space and Power Efficiency team’s scope covers both Google-owned sites and leased colocation space. (Google careers, undated, read 2026-10-15)
FactCustom silicon sold three ways: TPUs run in Google’s own cloud, are placed in customers’ own data centers, and are offered through a neocloud set up with Blackstone; for Anthropic, Google-built TPUs are also supplied through Broadcom. (Investing.com, 2026-09-08) (StockAnalysis, 2026-09-08) (Google Cloud, 2026-04-06)
FactPartner GPU capacity as a bridge: rather than wait for its own halls, Google contracted SpaceX for roughly 110,000 NVIDIA GPUs from October 2026, with termination rights if delivery slips. (SpaceX via SEC EDGAR, 2026-06-05) (CNBC, 2026-07-22)
FactPower arranged deal by deal: long-term nuclear agreements (Constellation, Duane Arnold, Loviisa), a 20-year fuel-cell agreement with AEP, utility-built solar and batteries in Indiana, and a stated policy in Alabama of paying 100% of its own power and infrastructure costs. (Google Cloud, 2026-10-06) (NextEra Energy, 2025-10-28) (Google, 2026-09-09) (ESG Today, 2026-10-15) (Indiana Public Media, 2026-10-12) (Google, 2026-06-15)
FactPlanning discipline inside the fleet: one plan of record for TPU and GPU demand across Alphabet, plus power-oversubscription and space-release methods to get more usable capacity out of buildings that already exist. (Google careers, undated, read 2026-10-15) (Google careers, undated, read 2026-10-15)

Latest updates, as of October 2026

2026-10-15AEP announced a $4 billion, 702-megawatt fuel-cell power plant near Cheyenne, Wyoming under a 20-year agreement with Google, with full operation by the end of 2028. (ESG Today, 2026-10-15)
2026-10-12Indiana regulators approved AES Indiana’s plan to serve a Google data center in Monrovia of at least 390 megawatts, with about $1.3 billion of grid, solar and battery investment. (Indiana Public Media, 2026-10-12)
2026-10-06Google and Constellation announced a 20-year agreement for 890 megawatts of new nuclear capacity on the PJM grid, with the first uprate expected by 2028. (Google Cloud, 2026-10-06)
2026-10-06Finnish authorities ordered Google’s subsidiary to pause environment-altering work at the Muhos and Kajaani sites by October 23, 2026, pending environmental assessments; Google said it would comply. (Helsinki Times, 2026-10-06)
2026-09-09Announced a 13 billion euro investment in Finland over two years, including a 22-year nuclear life-extension agreement and a 94-megawatt battery, with the main construction phase in 2027 and 2028. (Google, 2026-09-09)
2026-09-08At a Goldman Sachs conference, Google Cloud CEO Thomas Kurian said the TPU business is more than twice the size of the next hyperscaler’s accelerator business. (Investing.com, 2026-09-08)

Who leads it

Thomas Kurian, Chief Executive Officer, Google Cloud — runs Google Cloud, which sells TPU and GPU capacity to outside customers (Google Cloud, undated, read 2026-10-15)
Amin Vahdat, Chief Technologist for AI Infrastructure — technical lead for Google’s AI infrastructure (Google Cloud, undated, read 2026-10-15)
Bikash Koley, Vice President, Google Global Infrastructure — Google’s global infrastructure; announced the Finland build (Google, 2026-09-09)
Amanda Peterson Corio, Global head of energy and power, Google — energy and power supply for Google’s data centers; quoted on the Constellation nuclear deal (Google Cloud, 2026-10-06)

The roles that do this work there

Technical Program Manager, Data Center Capacity Delivery
Runs multi-gigawatt data-center build portfolios from site selection through construction, commissioning and handover, including late-stage GPU, TPU and liquid-cooling insertions.
Reston, VA; Austin, TX; Sunnyvale, CA and eight other U.S. cities · remote policy not stated · $192,000 – $278,000 + 20% bonus target + equity + benefits (U.S. base salary range) · posted not shown; applications open until at least 2026-10-26
Leadership Technical Program Manager II, Data Center Demand and Allocation Planning, Machine Learning
Leads the team that owns the Alphabet-wide TPU and GPU data-center plan of record and decides capacity trade-offs between demand and physical supply.
San Francisco, CA; Austin, TX; Kirkland, WA; Sunnyvale, CA · remote policy not stated · $192,000 – $278,000 + 20% bonus target + equity + benefits (U.S. base salary range) · posted not shown
Senior Technical Program Manager I, Capacity Planning, Google Data Centers
Reconciles usable power and floor space across the global fleet, runs power-oversubscription and space-release strategies, and signs off on space, power and cooling feasibility.
Sunnyvale, CA · remote policy not stated · $192,000 – $278,000 + 20% bonus target + equity + benefits (U.S. base salary range) · posted not shown
Technical Program Manager III, Capacity Planning, Google Data Centers
Manages cluster power reservations against machine demand to prevent deployment shortfalls and resolves space, power and network blockers with site teams.
Atlanta, GA · remote policy not stated · $163,000 – $236,000 + 15% bonus target + equity + benefits (U.S. base salary range) · posted not shown
Technical Program Manager III, Data Center Infrastructure, Technical Infrastructure
Represents data-center operations in new-product programs for power, cooling and server-hall infrastructure so new hardware is ready to operate on arrival.
Sunnyvale, CA; Lithia Springs, GA and 31 more · not stated; up to 20% travel for factory witness testing and site-readiness validation · $163,000 – $236,000 + 15% bonus target + equity + benefits (U.S. base salary range) · posted not shown
Technical Program Manager II, Capacity Constraints Management, Cloud Infrastructure
Plans accelerator deployments in colocation data centers, managing power, space, cooling and network constraints and the floor layouts new products need.
Atlanta, GA · remote policy not stated · $138,000 – $197,000 + 15% bonus target + equity + benefits (U.S. base salary range) · posted not shown

What it likely means, and what to watch

InferenceRaising the spending range because capacity is being delivered faster, while also renting third-party capacity as a bridge, suggests that in mid-2026 the binding constraint for Google was getting capacity live, not getting approval to spend.
InferenceA team whose stated purpose is to unlock stranded power and floor space implies Google believes meaningful capacity can be recovered inside buildings it already owns or leases, without waiting for new construction.
InferenceWith many power deals landing between 2028 and 2029, the multi-year gap between buying hardware and energizing it is the schedule risk capacity teams have to manage.

QuestionDid SpaceX meet the September 30, 2026 GPU delivery deadline, and will Google keep the $920 million-a-month contract past the December 31, 2026 point at which either side can give notice?
QuestionHow much does the Finnish pause at Muhos and Kajaani delay the 2027–2028 construction phase of the 13 billion euro plan?
QuestionWhen do the eighth-generation TPUs become generally available, and does Anthropic’s multi-gigawatt TPU capacity begin arriving on schedule in 2027?

Microsoft (Azure)

A cloud whose growth rate, by its own CFO’s account, moves with when capacity lands.

Why this problem matters to them

Microsoft spent $41 billion on capital equipment in a single quarter, and about two-thirds of that went to short-lived parts such as GPUs and CPUs. Those chips earn nothing until they sit in a finished building with power, cooling and networking and are handed to a customer. Microsoft’s own finance chief says customer demand still exceeds the capacity it can offer, and that Azure’s growth rate moves with when capacity lands. So for Microsoft, the gap between a GPU it has paid for and a GPU a customer can use is the gap between money spent and money earned, and turning purchases into ready capacity faster is how it keeps its promise to roughly double capacity in two years. (Microsoft Investor Relations, 2026-07-29)

It is like an airline that has bought a fleet of new planes but is still waiting on runways, fuel and crews. The planes are an expense on day one; they become revenue only when they fly with passengers.

Show what Microsoft (Azure) is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactBringing capacity online quickly: 31 new datacenters in the June 2026 quarter (88 for the fiscal year), another gigawatt in the quarter, and a goal of roughly doubling total capacity in two years. The first Fairwater building in Wisconsin went fully operational in June 2026, with a second due in 2028. (Microsoft Investor Relations, 2026-07-29) (Microsoft, 2026-06-23)
FactSpending heavily and guiding higher: $41 billion of capital spending in the June 2026 quarter, more than $50 billion expected in the next, and fiscal 2027 spending expected to grow; cash additions to property and equipment reached $115.9 billion for fiscal 2026. (Microsoft Investor Relations, 2026-07-29) (Microsoft Investor Relations, 2026-07-29)
FactBuilding its own chips to get more work out of each megawatt: Maia 200 is running production workloads such as Copilot, and Maia 300, announced October 12, 2026, is slated for multi-gigawatt deployment from 2027 as the primary inference system. (Microsoft, 2026-10-12) (Microsoft Investor Relations, 2026-07-29)
FactRenting ready capacity from specialist neocloud providers while its own sites are built: about 200,000 GB300 GPUs from Nscale across Texas, Portugal, the UK and Norway; a $9.7 billion, five-year IREN deal in Texas; a Nebius contract from New Jersey; a Lambda deal for tens of thousands of GPUs; and a 900-megawatt Crusoe campus in Abilene. (Nscale, 2025-10-15) (TechCrunch, 2025-11-03) (Nebius, 2025-09-08) (Lambda, 2025-11-03) (Crusoe, 2026-03-27)
FactHiring for readiness, not only construction: October 2026 postings for a principal program manager in the Integrated Planning & Execution team, another for AI Supercomputer Operations, a GPU Deployment Execution program manager, and a September director role to run its neocloud and GPU-as-a-Service sourcing strategy. (Microsoft careers, 2026-10-14) (Microsoft careers, 2026-10-12) (Microsoft careers, 2026-10-05) (Microsoft careers, 2026-09-21)

How they do it

FactOwned hyperscale AI campuses designed as one machine: the Fairwater sites in Wisconsin and Atlanta use roughly 140-kilowatt racks with up to 72 Blackwell GPUs each and closed-loop liquid cooling, networked so hundreds of thousands of GPUs can act as a single supercomputer. (Microsoft, 2026-06-23) (Microsoft, 2025-11-12)
FactNew owned campuses with power arranged up front: the Pecos, Texas campus is sized at about 2 gigawatts with a natural-gas facility operating at launch; the Middle East plan commits more than $10 billion through 2030. (Microsoft, 2026-06-22) (Microsoft, 2026-09-23)
FactPartner and neocloud capacity delivered in phases: Nscale from the first quarter of 2026 in Portugal and the third quarter in Texas, IREN through 2026, Nebius from late 2025, Lambda, and Crusoe’s Abilene campus from mid-2027. (Nscale, 2025-10-15) (TechCrunch, 2025-11-03) (Nebius, 2025-09-08) (Lambda, 2025-11-03) (Crusoe, 2026-03-27)
FactCustom silicon to raise useful output per unit of power: Maia 300 is aimed at inference, where Microsoft says tokens per megawatt matter most; Microsoft says it modernizes its fleet with its own silicon alongside the latest NVIDIA and AMD hardware. (Microsoft, 2026-10-12) (Microsoft Investor Relations, 2026-07-29)
FactFinance and organization levers: datacenter useful lives extended from 15 to 25 years, a shift toward operating leases that changes how spending is reported but not the spending itself, and a dedicated AI Capacity Business Development function to decide build, buy, partner or invest. (Microsoft Investor Relations, 2026-07-29) (Microsoft careers, 2026-09-21)

Latest updates, as of October 2026

2026-10-14Posted a remote Principal Technical Program Manager – Datacenter Infrastructure role whose stated job is global planning to maximize capacity efficiency across the datacenter portfolio. (Microsoft careers, 2026-10-14)
2026-10-12Announced Azure Maia 300 and said it will scale Maia to multi-gigawatt deployments from 2027 as its primary inference system; Maia 200 is already serving Copilot. (Microsoft, 2026-10-12)
2026-10-12Posted a remote Principal Technical Program Manager, AI Supercomputer Operations role covering availability, repair and recovery, and readiness for its largest AI supercomputers. (Microsoft careers, 2026-10-12)
2026-10-05Posted a remote GPU Deployment Execution Program Manager role with 50–75% travel and a printed pay range. (Microsoft careers, 2026-10-05)
2026-09-23Brad Smith announced more than $10 billion of cloud and AI infrastructure spending in the Middle East through 2030. (Microsoft, 2026-09-23)

Who leads it

Noelle Walsh, President, Cloud Operations and Innovation — leads the organization that plans, builds and operates Microsoft’s datacenters; announced the Pecos campus (Microsoft, 2026-06-22)
Scott Guthrie, Executive Vice President, Cloud + AI — leads the group that includes Azure, the Fairwater AI datacenters and the Maia chip program (Microsoft, 2026-10-12)
Amy Hood, Executive Vice President and Chief Financial Officer — sets and explains capital-spending guidance and capacity-versus-demand commentary (Microsoft Investor Relations, 2026-07-29)
Jon Tinter, President, Business Development and Ventures — the Microsoft executive quoted on the Nscale capacity agreement (Nscale, 2025-10-15)

The roles that do this work there

Principal Technical Program Manager – Datacenter Infrastructure
Leads multi-year programs that maximize how efficiently Microsoft’s global datacenter capacity is used while keeping sites safe, secure and available.
United States, multiple locations · 0 days a week in office; remote · USD $142,800 – $274,800 per year (USD $188,000 – $304,200 in the San Francisco Bay Area and New York City) (typical base pay range) · posted 2026-10-14
Principal Technical Program Manager, AI Supercomputer Operations
Runs day-to-day operations of Microsoft’s largest AI training supercomputers: availability, incident response, repair and recovery, planned maintenance and operational readiness.
United States, multiple locations · 0 days a week in office; remote · USD $142,800 – $274,800 per year (USD $188,000 – $304,200 in the San Francisco Bay Area and New York City) (typical base pay range) · posted 2026-10-12
GPU Deployment Execution Program Manager
Program-manages the execution of GPU deployments across Microsoft datacenters.
Redmond, WA and four more · 0 days a week in office; remote, with 50–75% travel · USD $102,600 – $202,800 per year (USD $137,500 – $225,800 in the San Francisco Bay Area and New York City) (typical base pay range) · posted 2026-10-05
Director – AI Capacity Business Development
Leads Microsoft’s external strategy for sourcing AI capacity, including partnerships with neocloud and GPU-as-a-Service providers and build, buy, partner or invest decisions.
Redmond, WA and one more · 3 days a week in office · USD $130,900 – $251,900 per year (USD $165,600 – $272,300 in the San Francisco Bay Area and New York City) (typical base pay range) · posted 2026-09-21

What it likely means, and what to watch

InferenceMicrosoft’s own words, that demand exceeds supply and Azure growth depends on capacity timing, suggest the binding constraint is how fast purchased hardware becomes usable capacity, not the purchase budget.
InferenceCommitting Maia to multi-gigawatt scale and framing its benefit as tokens per megawatt implies Microsoft treats power, not chips, as the scarce unit.
InferenceA September 2026 director posting that names neoclouds and GPU-as-a-Service as a standing strategy area suggests third-party capacity is a lasting part of the plan rather than a temporary stopgap.

QuestionWill Microsoft give a dollar figure for fiscal 2027 capital spending on its late-October 2026 call, and how much of it is short-lived GPU spend versus buildings and leases?
QuestionWhen does Microsoft say Azure stops being capacity-constrained, and does the “double capacity in two years” goal still hold?
QuestionDo the phased partner deliveries and the second Wisconsin Fairwater building, due in 2028, arrive on schedule?

Amazon / AWS

The largest spender, selling gigawatts of its own chips years ahead and telling investors it still will not have enough capacity.

Why this problem matters to them

Amazon told investors it now expects to spend about $220 billion in 2026, and that even then it will not have enough capacity to meet demand in 2026 or 2027. It has already sold multi-gigawatt blocks of that future capacity: up to 5 gigawatts to Anthropic, with nearly 1 gigawatt due online by the end of 2026, and about 2 gigawatts of Trainium to OpenAI. Amazon’s own filing says those contracts carry obligations tied to how its chips perform. That likely makes the time between a chip arriving and a chip performing a commercial question, not only an engineering one, which is why planning, delivery and commissioning matter so much to how the $220 billion turns into revenue. (Amazon, via Investing.com, 2026-07-30) (Amazon, via The Motley Fool, 2026-08-07) (Anthropic, 2026-04-20) (Amazon via SEC EDGAR, 2026-02-27) (Amazon via SEC EDGAR, 2026-07-30)

It is like a city that buys a thousand new buses but has not yet built the depots, hired the drivers or signed the fuel contracts. The buses are paid for, but nobody can ride until the rest is finished.

Show what Amazon / AWS is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactSpending at record scale: 2026 cash capital spending guidance was raised to about $220 billion on the July 30, 2026 call, after $98.4 billion of property-and-equipment purchases in the first half of the year, and Amazon says it aims to double its 2025 power capacity by the end of 2027. (Amazon, via Investing.com, 2026-07-30) (Amazon via SEC EDGAR, 2026-07-30) (Amazon, via The Motley Fool, 2026-08-07) (Amazon Investor Relations, 2026-07-30)
FactLocking in anchor customers for its own chips: Anthropic committed to more than $100 billion over ten years and up to 5 gigawatts of capacity; OpenAI expanded its AWS agreement by $100 billion over 8 years and committed to roughly 2 gigawatts of Trainium. Amazon’s 10-Q says both arrangements include obligations tied to AWS chip performance. (Anthropic, 2026-04-20) (Amazon, 2026-04, page undated) (Amazon via SEC EDGAR, 2026-02-27) (Amazon via SEC EDGAR, 2026-07-30)
FactBuilding named campuses in phases: a third campus in Shreveport that lifts planned northwest Louisiana investment to $18 billion, and a 13-building campus near Boling, Texas. (Amazon, 2026-08-18) (Amazon, 2026-09-11)
FactArranging the power and water those buildings need: a revised nuclear power agreement with Talen for up to 1,920 megawatts through 2042 took effect in April 2026; a framework with Indiana utility NIPSCO, described in a September 2026 update, to add up to 3 gigawatts of grid capacity; more than 10 gigawatts of energy supply in Texas; and a $480 million reclaimed-water project in Virginia. (Talen Energy via SEC EDGAR, 2026-08) (Amazon, 2026-09-25)
FactHiring the people who turn buildings into usable capacity: planners who align demand with supply, power and capital plans; capacity-delivery planners working a 0-to-24-month horizon; delivery managers who install server racks; and commissioning engineers who test electrical and mechanical systems, with thirteen Texas commissioning openings posted on October 14, 2026 alone. (Amazon careers, 2026-10-14) (Amazon careers, 2026-07-08) (Amazon careers, 2026-09-21) (Amazon careers, 2026-10-14)
FactTreating local approval as part of readiness: in October 2026 AWS CEO Matt Garman announced Built Together, more than $1 billion over five years for data-center communities, annual publication of energy and water use, and an end to nondisclosure agreements with government agencies, citing over 100 moratoriums under consideration. (Amazon, 2026-10, page undated) (Semafor, 2026-10-02)

How they do it

FactCustom silicon as the main lever: Trainium2 in Project Rainier (nearly half a million chips per Amazon; over one million in use per Anthropic), Trainium3 UltraServers of up to 144 chips each, generally available since December 2025, and Trainium4 expected to begin delivery in 2027. A chip is a processor; an UltraServer is many chips wired together as one machine. (Amazon (press release), 2025-12-02) (Amazon, 2026-04, page undated) (Anthropic, 2026-04-20) (Amazon via SEC EDGAR, 2026-02-27)
FactOwned data centers built in waves rather than all at once: Louisiana is three separate campuses, and Boling is 13 buildings. Each wave is planned, delivered and commissioned before it counts as capacity. (Amazon, 2026-08-18) (Amazon, 2026-09-11)
FactPower through long-term contracts and utility deals: a 1,920-megawatt nuclear supply agreement with Talen through 2042, plus utility frameworks in Indiana where Amazon pays grid-upgrade fees. (Talen Energy via SEC EDGAR, 2026-08) (Amazon, 2026-09-25)
FactA planning-and-delivery organization: AWS Infrastructure Services owns design, planning, delivery and operation of all AWS infrastructure; a planning cycle reconciles demand, supply, power and capital; capacity-delivery planners work a 0-to-24-month horizon; delivery managers install racks; commissioning engineers run factory witness tests and on-site functional and integrated tests. (Amazon careers, 2026-10-12) (Amazon careers, 2026-10-14) (Amazon careers, 2026-07-08) (Amazon careers, 2026-09-21) (Amazon careers, 2026-10-14)
FactInvestment-for-commitment deals with AI labs: equity in Anthropic ($5 billion plus up to $20 billion more) and OpenAI ($50 billion) paired with $100 billion-plus AWS spending commitments from each lab. (Amazon, 2026-04, page undated) (Amazon via SEC EDGAR, 2026-07-30)

Latest updates, as of October 2026

2026-10-14Posted thirteen Commissioning Engineer openings across Dallas, Houston and Coyanosa, Texas in one day, plus a Senior S&OP Planner role in Seattle focused on aligning demand with supply, power and capital plans. (Amazon careers, 2026-10-14)
2026-10-02AWS CEO Matt Garman announced Built Together, more than $1 billion over five years for data-center communities, an Amazon Data Center Commitment, annual publication of energy and water use, and an end to nondisclosure agreements with government agencies. (Amazon, 2026-10, page undated)
2026-09-25Amazon’s update of this date describes a framework developed with Indiana utility NIPSCO to add up to 3 gigawatts of new capacity to the Indiana grid, with Amazon paying fees for existing lines and needed upgrades. (Amazon, 2026-09-25)
2026-09-11Described a planned 13-building data-center campus near Boling, Texas, with site preparation under way, and said it has invested in more than 10 gigawatts of energy supply in Texas. (Amazon, 2026-09-11)
2026-09-10Announced with Spotsylvania County, Virginia a $480 million water-reuse project that will pipe reclaimed water to cool Amazon’s data centers. (Amazon, 2026-09-10)
2026-08-18Raised its planned investment in northwest Louisiana to $18 billion with a third planned campus in Shreveport. (Amazon, 2026-08-18)

Who leads it

Matt Garman, CEO, Amazon Web Services — runs AWS; author of the October 2026 data-center commitment post (Amazon, 2026-10, page undated)
Andy Jassy, President and CEO, Amazon — sets company capital spending and the goal of doubling 2025 power capacity by the end of 2027 (Amazon via SEC EDGAR, 2026-02-27)

The roles that do this work there

Senior S&OP Planner, Infrastructure Capacity & S&OP Planning
Owns the monthly and quarterly planning cycle that lines up AWS demand signals with supply, power and capital plans and runs the executive planning forums.
Seattle, WA · remote policy not stated · 148,700.00 – 201,200.00 USD annually (Seattle) (base salary; sign-on and restricted stock units additional) · posted 2026-10-14
Senior Technical Infrastructure Program Manager, Infrastructure Capacity Management
Runs supply-chain and capacity-delivery programs end to end so AWS data centers get capacity on time.
Seattle, WA · remote policy not stated · 148,700.00 – 201,200.00 USD annually (Seattle) (base salary; sign-on and restricted stock units additional) · posted 2026-10-12
Tech Infra Program Manager, Capacity Delivery Planning
Plans data-center capacity for each AWS region over a 0-to-24-month horizon, clears racks-on-hold and blocked orders, and manages projects to hit power-usage targets.
Herndon, VA · remote policy not stated · 148,700.00 – 201,200.00 USD annually (Herndon) (base salary; sign-on and restricted stock units additional) · posted 2026-07-08
Technical Infrastructure Program Manager, Data Center Planning Delivery
Leads data-center build projects from design feasibility through budget, implementation, testing and deployment.
Atlanta, GA or Charlotte, NC (alternates Herndon, VA and Seattle, WA) · remote policy not stated · 127,100.00 – 171,900.00 USD annually (Atlanta) (base salary; sign-on and restricted stock units additional) · posted 2026-08-25
Data Center Infrastructure Delivery Manager, DCC Communities
Manages the technicians who install and build server-rack capacity at the New Carlisle, Indiana campus.
New Carlisle, IN · remote policy not stated · 78,400.00 – 137,300.00 USD annually (New Carlisle) (base salary; sign-on and restricted stock units additional) · posted 2026-09-21
Infrastructure Delivery Cluster Manager
Leads network capacity delivery for a cluster of data centers, managing inbound equipment from receiving through deployment with a zero-defects goal.
Frederick, MD · remote policy not stated · 145,400.00 – 254,500.00 USD annually (Frederick) (base salary; sign-on and restricted stock units additional) · posted 2026-10-08
Commissioning Engineer, AWS Commissioning, AMER-West
Tests and validates a new data center’s electrical and mechanical systems, leading factory witness tests and managing third-party vendors through on-site functional and integrated testing.
Dallas, TX (identical postings in Houston and Coyanosa, TX) · on-site testing at data-center sites · 116,800.00 – 160,000.00 USD annually (Dallas) (base salary; sign-on and restricted stock units additional) · posted 2026-10-14

What it likely means, and what to watch

InferenceJassy said higher memory costs pushed the spending figure up, which suggests that in 2026 the cost and supply of components is as much a readiness constraint as construction.
InferenceBecause the 10-Q says the Anthropic and OpenAI arrangements include obligations tied to AWS chip performance, AWS likely carries delivery risk until chips are installed, powered and performing, which makes commissioning and capacity-delivery work directly revenue-relevant.
InferenceThirteen commissioning-engineer openings posted in one day across three Texas locations likely signal Texas campuses moving from construction to turn-on during 2027.

QuestionWill nearly 1 gigawatt of Trainium2 and Trainium3 capacity for Anthropic be online by the end of 2026, as Anthropic’s April 2026 post expected?
QuestionOn the late-October 2026 earnings call, will Amazon keep the roughly $220 billion figure and the goal of doubling 2025 power capacity by the end of 2027, and will it report gigawatts actually brought online?
QuestionDo the 100-plus proposed moratoriums or utility proceedings slow any of the named campuses in Indiana, Texas, Louisiana, North Carolina or Virginia?

Oracle (OCI)

A cloud with $664 billion of contracted work, most of which becomes revenue only as campuses are built, powered and accepted.

Why this problem matters to them

Oracle has signed far more AI cloud business than it can serve today. Its backlog of contracted work was $664 billion at the end of August 2026, but Oracle expects only about 13% of it to become revenue in the next twelve months, because revenue arrives only as sites are built, powered, filled with GPUs and accepted by customers. That is why Oracle tied its 30% revenue growth last quarter to delivering 850 megawatts of new capacity. It has also signed $288 billion of future data-center leases that have not started yet. Each site that waits on power, a pipeline or a permit keeps contracted money in backlog instead of revenue. (Oracle, 2026-09-10) (Oracle via SEC EDGAR, 2026-09-11)

It is like a restaurant that has pre-sold months of dinners and bought all its ovens, but the gas line to the building is not connected yet, so it trucks in propane to light the first burners.

Show what Oracle (OCI) is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactDelivering capacity quickly: 850 megawatts and more than 300,000 GPUs delivered in the June–August 2026 quarter, with the Abilene, Texas campus at 618 megawatts across six of eight buildings and customer acceptance of new systems down to 24 hours. (Oracle, 2026-09-10) (Oracle, via The Motley Fool, 2026-09-11)
FactFunding the build: fiscal 2027 capital spending guided at $90 billion to $95 billion, and more than $30 billion of new AI contracts structured with prepayment or customer-supplied hardware. (Oracle, via The Motley Fool, 2026-09-11) (Oracle, 2026-09-10)
FactBringing the next gigawatt-class campuses online: Shackelford County, Texas; Project Jupiter in New Mexico; The Barn in Michigan; and Project Lighthouse in Wisconsin. None of these delivered capacity in the June–August quarter. (Oracle, via The Motley Fool, 2026-09-11) (Oracle, 2026-09-29) (Oracle, 2026-09-28) (Oracle, 2026-10-02)
FactArranging power alongside the buildings: Oracle committed to subscribe to Point Beach nuclear output in Wisconsin, and says Project Jupiter in New Mexico remains on its planned schedule after flooding reached the campus. (Oracle, 2026-10-02) (Oracle, 2026-09-30)
FactHiring the people who turn sites into usable GPU capacity: a principal program manager for data centers, a director of data-center infrastructure delivery, a capacity planner, a facilities development manager who leads commissioning, and on-site power-generation coordinators, all posted between August 20 and October 15, 2026. (Oracle careers, 2026-10-12) (Oracle careers, 2026-08-31) (Oracle careers, 2026-08-20) (Oracle careers, 2026-10-15) (Oracle careers, 2026-10-13)

How they do it

FactLeased, developer-built campuses rather than fully owned ones: $288 billion of future data-center lease commitments with 15-to-19-year terms, and postings that describe leading construction of data centers by colocation providers. (Oracle via SEC EDGAR, 2026-09-11) (Oracle careers, 2026-10-15)
FactGigawatt-scale campus and cluster design built around GPU networking: the Zettascale10 design, first deployed at Abilene, targets clusters of up to 800,000 NVIDIA GPUs within a two-kilometre radius. (Oracle, 2025-10-14) (Oracle, via The Motley Fool, 2026-09-11)
FactPower treated as its own workstream: on-site generation with dedicated operations coordinators and a nuclear subscription in Wisconsin. (Oracle careers, 2026-10-13) (Oracle, 2026-10-02)
FactContract structures that shift cash timing: the new AI contracts use prepayment or customer-supplied hardware, so Oracle says they need no additional cash from it even though capital spending still occurs. (Oracle, 2026-09-10) (Oracle, via The Motley Fool, 2026-09-11)
FactCapacity delivery run as a program discipline: a program manager owns the path from site selection through construction, commissioning, fit-out, network and server readiness to handover, and capacity planners balance supply and demand across every region. (Oracle careers, 2026-10-12) (Oracle careers, 2026-08-31) (Oracle careers, 2026-08-20)

Latest updates, as of October 2026

2026-10-12Posted a Principal Technical Program Manager, Data Centers role whose stated job is converting a site into live GPU capacity, from site selection through commissioning and handover. (Oracle careers, 2026-10-12)
2026-10-02Committed to subscribe to part of the Point Beach nuclear plant’s output in Wisconsin, tied to Project Lighthouse and pending regulatory approval. (Oracle, 2026-10-02)
2026-09-10First-quarter fiscal 2027 results: revenue up 30% to $19.3 billion on 850 megawatts of new capacity, cloud infrastructure revenue up 121%, and $664 billion of contracted backlog. (Oracle, 2026-09-10)

Who leads it

Clay Magouyrk, Chief Executive Officer — co-CEO; previously President of Oracle Cloud Infrastructure, and the executive who reports capacity delivery on earnings calls (Oracle, 2025-09-22)
Mike Sicilia, Chief Executive Officer — co-CEO; formerly President, Oracle Industries (Oracle, 2025-09-22)
Mahesh Thiagarajan, Executive Vice President and Head of Oracle Cloud Infrastructure — runs OCI; quoted on the Point Beach power commitment for the Wisconsin data center (Oracle, undated, read 2026-10-15)
Hilary Maxson, Chief Financial Officer — capital-spending guidance and how new backlog flows into capex and revenue (Oracle, undated, read 2026-10-15)

The roles that do this work there

Principal Technical Program Manager, Data Centers
Leads hyperscale capacity-delivery programs from site selection through design, construction, commissioning, fit-out, network and server readiness to customer handover.
United States · remote policy not stated · $102,300 to $209,500 per annum (U.S. hiring range) · posted 2026-10-12
Principal Data Center Capacity Planner
Manages supply and demand across OCI’s fleet of regions and scopes large customer and next-generation AI cluster requirements with sales, finance and engineering.
Nashville, TN · remote policy not stated · $90,100 to $209,500 per annum (U.S. hiring range) · posted 2026-08-20
Director, Data Center Infrastructure Delivery
Leads the delivery lifecycle for large-scale data-center capacity, from site and colocation engagement through construction, commissioning, operational readiness and handoff.
United States (ideally Texas) · remote policy not stated; ideally based in Texas · $146,300 to $306,400 per annum (U.S. hiring range) · posted 2026-08-31
Principal Data Center Facilities Development Manager
Leads construction of data centers by colocation providers, oversees quality control, and leads startup and commissioning.
United States · remote policy not stated · $102,300 to $209,500 per annum (U.S. hiring range) · posted 2026-10-15
Operations Coordinator – Data Center Power Generation
Coordinates the real-time readiness of on-site power-generation assets, short-term load forecasts and fuel nominations so generation is available to meet data-center demand.
Saline, MI; Port Washington, WI; or New Mexico · full time on site at one of the three data-center sites · $83,000 to $166,100 per annum (U.S. hiring range) · posted 2026-10-13
Technical Program Manager, Hardware NPI & Server Platforms
Leads new-product-introduction programs for next-generation GPU and CPU server platforms from requirements through qualification, manufacturing readiness and production ramp.
Santa Clara, CA (also Seattle, Houston, Burlington MA) · remote policy not stated · $90,100 to $209,500 per annum (U.S. hiring range) · posted 2026-10-13

What it likely means, and what to watch

InferenceOracle’s growth now looks paced by physical delivery rather than sales: its own schedule converts the $664 billion backlog over several years, and management says new backlog will not affect revenue until fiscal 2028 or later.
InferencePower, not GPUs, looks like a gating item: Oracle is hiring on-site power-generation coordinators, and its next gigawatt-class campuses delivered no capacity in the June–August quarter.
InferencePrepayment and customer-supplied-hardware structures move cash timing to customers and suppliers, but the readiness work of construction, commissioning, acceptance and utilization stays with Oracle.

QuestionDoes Project Jupiter stay on its planned schedule after flooding reached the campus, as Oracle said on September 30, 2026?
QuestionHow many megawatts from Shackelford, New Mexico, Wisconsin and Michigan are delivered in the quarter ending November 2026, and does the 24-hour customer acceptance seen at Abilene hold at new sites?
QuestionDoes GPU utilization stay near 97.9% as 850-megawatt quarters continue, or does supply start to catch up with demand?

CoreWeave

A GPU cloud that reports its progress in megawatts: how much power is under contract against how much is actually live.

Why this problem matters to them

CoreWeave sells one thing: time on GPU clusters that live in data centers. Customers have signed contracts worth about $104 billion, but CoreWeave earns that money only as each cluster is built, powered, tested and handed over. In August 2026 it had about 4.2 gigawatts of power under contract and 1.5 gigawatts actually live, so most of its future revenue still depends on turning signed power into working machines. Spending runs ahead of income: the company guided to $35 billion to $39 billion of capital spending in 2026, and more than half its backlog is tied to contracts where delivery has already started. If a site opens late, the hardware bought for it waits without earning. (CoreWeave, 2026-08-11) (CoreWeave, 2026-09-17) (CoreWeave, 2026-08-11)

It is like a hotel chain that has pre-sold rooms for the next five years while several of its hotels are still being wired. The bookings are real, but the money arrives only as each floor passes inspection and guests get their keys.

Show what CoreWeave is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactBringing power live fast: nearly 500 megawatts of active power added in the second quarter of 2026, more than 300 of it in June, to reach 1.5 gigawatts, with more than 1.85 gigawatts now expected by year-end. (CoreWeave, 2026-08-11) (CoreWeave, 2026-08-11)
FactLocking in future power: about 4.2 gigawatts contracted by August 11, 2026, a first self-built site expected online later in 2026, and a stated target of at least 8 gigawatts of active power by 2030. (CoreWeave, 2026-09-17) (CoreWeave, 2026-08-11)
FactExpanding outside the United States: more than a gigawatt contracted abroad, including 360 megawatts in Indonesia due in roughly 18 months, and a 240-megawatt campus in Navi Mumbai whose first phase is expected in mid-2028. (CoreWeave, 2026-08-11) (CoreWeave, 2026-10-06)
FactSpending and financing the buildout: $9.4 billion of capital spending in the second quarter, $11.5 billion to $13.5 billion guided for the third, $35 billion to $39 billion for the year, and a September investor deck for a convertible notes offering. (CoreWeave, 2026-08-11) (CoreWeave via SEC EDGAR, 2026-09)
FactHiring a dedicated delivery organization: on October 15, 2026 the job board carried Capacity Delivery and Fleet Delivery program managers, an expansion PMO director, a data-center operations program manager and a commissioning and quality manager, plus a Partner Network so third-party tools are tested under production conditions. (CoreWeave careers, 2026-10-15) (CoreWeave, 2026-09-30)

How they do it

FactMostly partner-built and leased data centers: CoreWeave’s own investor deck shows partners taking 12 to 18 months to develop and build a site, then CoreWeave spending about 1 to 2 months installing and testing GPUs; examples include the AdaniConneX campus in India and anchoring the first phase of Poolside’s West Texas campus. (CoreWeave via SEC EDGAR, 2026-09) (CoreWeave, 2026-10-06) (CoreWeave, 2025-10-15)
FactStarting to self-build: the first self-built site is expected online later in 2026, and powered land, site options and letters of intent give it further potential beyond what is already contracted. (CoreWeave, 2026-08-11)
FactA deep NVIDIA tie-in: NVIDIA invested $2 billion in January 2026 to help accelerate more than 5 gigawatts of AI factories by 2030, including help buying land, power and building shells, and Meta’s $21 billion expanded agreement includes some of the first Vera Rubin deployments. (CoreWeave via SEC EDGAR, 2026-01-26) (CoreWeave, 2026-04-09)
FactSupply chain and operations under one executive: the COO’s stated remit covers supply chain, procurement, capacity planning, IT engineering, data-center operations and international expansion, and the company calls the global supply chain complex. (CoreWeave, undated, read 2026-10-15) (CoreWeave, 2026-08-11)
FactContracts and money measured in megawatts: capital spending is front-loaded and funded with debt, customer prepayments and corporate capital, revenue ramps once a cluster is delivered, and new short contracts in the third quarter were priced at about $40 million per megawatt. (CoreWeave, 2026-08-11) (CoreWeave, 2026-09-17)

Latest updates, as of October 2026

2026-10-06Entered India with AdaniConneX: 240 megawatts across three buildings in Navi Mumbai, first phase expected in mid-2028. (CoreWeave, 2026-10-06)
2026-09-30Launched the CoreWeave Partner Network, listing partner integrations tested on its cloud under production conditions. (CoreWeave, 2026-09-30)
2026-09-17Said it signed three-to-six-month contracts at about $40 million per megawatt in the third quarter, and that contracted power reached about 4.2 gigawatts as of August 11. (CoreWeave, 2026-09-17)
2026-09Filed an investor presentation for a convertible notes offering listing 51 active data centers, about 4.2 gigawatts contracted and about 1.5 gigawatts active. (CoreWeave via SEC EDGAR, 2026-09)
2026-08-11Second-quarter results: backlog about $104 billion, active power up nearly 500 megawatts to 1.5 gigawatts, contracted power about 3.7 gigawatts. (CoreWeave, 2026-08-11)
2026-08-11On the call, said it now expects full-year 2026 capex of $35 billion to $39 billion and raised year-end active-power guidance to more than 1.85 gigawatts. (CoreWeave, 2026-08-11)

Who leads it

Michael Intrator, Co-founder, Chairman, President and CEO — runs the company and sets its power strategy, including the target of at least 8 gigawatts of active power by 2030 (CoreWeave, 2026-08-11)
Nitin Agrawal, Chief Financial Officer — capital spending, financing and guidance for the buildout (CoreWeave, 2026-08-11)
Sachin Jain, Chief Operating Officer — supply chain, procurement, capacity planning, IT engineering, data-center operations and international expansion (CoreWeave, undated, read 2026-10-15)
Brannin McBee, Co-founder and Chief Development Officer — business development, financing and strategic partnerships behind new sites and power (CoreWeave, undated, read 2026-10-15)

The roles that do this work there

Sr. Technical Program Manager, Capacity Delivery
Owns end-to-end delivery of data-center builds and capacity expansions, turning long-range capacity forecasts and customer requirements into fully commissioned infrastructure.
Livingston, NJ / New York, NY / Sunnyvale, CA / San Francisco, CA / Bellevue, WA · remote policy not stated; office hubs listed · $157,000 to $210,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-10-09
Staff Technical Program Manager, Capacity Delivery
Single-threaded leader for delivering critical data-center infrastructure from planning to operational delivery, and for the repeatable processes behind how the fleet gets built.
Livingston, NJ / New York, NY / Sunnyvale, CA / Bellevue, WA / San Francisco, CA · remote policy not stated; office hubs listed · $143,000 to $210,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-02-12, updated 2026-10-12
Senior Technical Program Manager – Fleet Delivery & Capacity Management
Leads delivery of the GPU and CPU fleet from infrastructure handover through provisioning, validation and acceptance, and surfaces gaps between requested and forecast availability.
New York, NY / Sunnyvale, CA / Bellevue, WA · remote policy not stated; office hubs listed · $182,000 to $242,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-09-16
Director, Expansion PMO, Data Center Planning
Owns program management of data-center expansion and supply strategy, aligning demand to supply across land and lease paths in new geographies.
US office hubs, with a global remit · based in US hubs; periodic international travel · $180,000 to $240,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-06-02
Data Center Operations Program Manager
Leads execution of data-center operations programs and clears bottlenecks in hardware logistics, rack integration and facility readiness.
Livingston, NJ / New York, NY / Sunnyvale, CA / San Francisco, CA / Bellevue, WA · remote policy not stated; office hubs listed · $134,000 to $179,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-06-16, updated 2026-10-14
Data Center Commissioning/Quality Manager
Oversees installation and testing of generators, UPS, electrical distribution and cooling so construction projects finish on time and to quality standards.
Must live near Northern Virginia, Dallas, Richmond VA, Livingston NJ or Atlanta · remote-eligible with 50% travel · $143,000 to $191,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-07-28
Technical Deployment Lead
Member of a travelling team that builds and deploys new sites across the U.S. and Canada, troubleshoots hardware and networks and runs quality audits.
Sunnyvale, CA / Bellevue, WA / Reno, NV / Phoenix, AZ · 100% on site at a West-coast data center, with up to 60% travel · $90,000 to $102,000 (base salary, plus discretionary bonus, equity and benefits) · posted 2026-08-03

What it likely means, and what to watch

InferenceThe gap between about 4.2 gigawatts contracted and about 1.5 gigawatts active implies roughly 2.7 gigawatts still in construction, commissioning, installation or acceptance; that in-flight work is what the Capacity Delivery, Fleet Delivery and Commissioning roles on the job board exist to push through.
InferenceManagement tied its higher 2026 spending to capacity it expects to deliver this year and said several customers want each GPU it brings online, which suggests delivery timing, not demand, is what paces revenue.
InferencePricing contracts per megawatt and reporting active against contracted power each quarter shows that power, not GPU count, is the unit CoreWeave itself uses to measure how much purchased capacity is ready to sell.

QuestionDoes active power reach the more-than-1.85-gigawatt year-end 2026 guide, and how much of the 4.2 gigawatts contracted becomes active during 2027?
QuestionDoes the first self-built site come online later in 2026 as stated, and what share of future capacity will be self-built rather than leased?
QuestionWho are the customers behind the more than $25 billion of early third-quarter commitments, and when do their deliveries start?

SpaceX / xAI

A rocket company that now owns an AI lab and rents its Colossus supercomputers to other labs by the month.

Why this problem matters to them

SpaceX now owns xAI, and it rents its Colossus computers to outside customers by the month. Some of those contracts carry delivery dates: the Google agreement gives Google remedies if the committed GPUs are not delivered by September 30, 2026, and either side can leave after December 31, 2026 with 90 days’ notice. The power side is also on a clock: 69 temporary gas turbines in Southaven must be removed by July 2027 while a permanent 1.2-gigawatt plant is built. (TechCrunch, 2026-06-05) (Quartz via Yahoo News, 2026-07-31)

It is like a restaurant that buys forty new ovens for a big catering contract but is still waiting on its permanent gas line. The ovens are paid for, the customer is waiting, and only the burners on the temporary supply can cook.

Show what SpaceX / xAI is doing, latest updates, who leads it and the roles →

What they are doing about it now

FactSelling capacity to other labs: Anthropic rents all of Colossus 1, more than 300 megawatts; Google pays $920 million a month from October 2026 for about 110,000 GPUs; Reflection AI pays $150 million a month for GB300 access at Colossus 2. (CNBC, 2026-05-06) (TechCrunch, 2026-06-05) (Quartz via Yahoo Finance, 2026-06-22)
FactSwapping temporary power for permanent power: under a July 2026 order, all 69 temporary turbines at Southaven must be removed by July 2027; by August 31, 2026 the company said 11 were powered down, with a 1.2-gigawatt plant of 41 permitted turbines being built. (CNBC, 2026-03-10) (Quartz via Yahoo News, 2026-07-31) (Magnolia Tribune, 2026-08-31)
FactHiring builders and commissioners: on October 15, 2026 the xAI board showed about 35 data-center and commissioning roles, almost all in Memphis and Southaven. (SpaceXSI careers, 2026-10-15) (SpaceXSI careers, 2026-08-18)
FactPlanning compute in orbit: SpaceX asked the FCC for up to 1 million orbital data-center satellites, revealed the AI1 satellite in June 2026 and announced a satellite factory in Bastrop, Texas. (Via Satellite, 2026-02-02) (Quartz, 2026-06-10)

How they do it

FactOwn, self-built data centers rather than cloud rentals: the company says it built the first Colossus in 122 days and doubled it in 92, and the Memphis-area complex is reported at about 2 million square feet across two sites. (SpaceXAI, undated, read 2026-10-15) (Fortune, 2026-07-19)
FactOn-site gas generation as the power bridge: temporary turbines powered the site while a permitted 41-turbine, 1.2-gigawatt plant is built; the temporary units must be gone by July 2027. (CNBC, 2026-03-10) (Quartz via Yahoo News, 2026-07-31) (Magnolia Tribune, 2026-08-31)
FactNVIDIA GPUs in its own clusters: the company’s Colossus page describes a first cluster of 200,000 NVIDIA H100 GPUs with a roadmap to 1 million. (SpaceXAI, undated, read 2026-10-15)
FactA dedicated compute supply-chain function: SpaceX postings seek managers to source GPUs, CPUs, storage, servers and racks for its training clusters, plus engineers who scale on-site power on a schedule set by compute demand. (SpaceX careers, 2026-07-24) (SpaceXSI careers, 2026-08-18)
FactA long-term bet on space-based compute: the FCC filing argues compute will be cheapest in orbit, and AI1 is designed to carry roughly one large GPU rack’s worth of power per satellite. (Via Satellite, 2026-02-02) (Quartz, 2026-06-10)

Latest updates, as of October 2026

2026-10-15The xAI job board, now under the SpaceXSI name, lists about 35 data-center and commissioning roles, almost all in Memphis and Southaven. (SpaceXSI careers, 2026-10-15)
2026-10-05Musk confirmed SpaceXAI will be renamed SpaceXSI; job postings already use the new name. (Euronews, 2026-10-05)
2026-08-31The company said it had powered down 11 of the 69 temporary turbines at Southaven, with the rest to come out over the next year. (Magnolia Tribune, 2026-08-31)

Who leads it

High Power Engineering and data-center construction and commissioning teams (SpaceXSI) — build, commission and scale power and data-center capacity at the Memphis and Southaven sites; no public source names their leaders (SpaceXSI careers, 2026-08-18)

The roles that do this work there

Program Manager, Data Center Construction – Memphis
Manages construction project managers delivering data-center builds from pre-construction through commissioning and handover.
Southaven, MS; Memphis, TN · primarily on site in the Memphis area, with travel to other U.S. project sites · pay not shown on posting · posted 2026-07-01
Commissioning Engineer (Electrical) – Memphis
Tests and starts up electrical systems at new data-center sites so purchased hardware can be powered safely.
Memphis, TN · on site, with travel to data-center construction sites across the U.S. · pay not shown on posting · posted 2026-10-02
Mechanical Commissioning Lead – Memphis
Leads commissioning of cooling and mechanical systems from buildout through startup.
Southaven, MS; Memphis, TN · on site, with extensive travel to construction sites · pay not shown on posting · posted 2026-05-05
Lead Engineer, Data Center – Memphis
Works with a small on-site team to scale high-capacity power systems on a schedule set by compute demand.
Memphis, TN · on site, with potential travel to other sites · pay not shown on posting · posted 2026-08-18
Manager, Supply Chain – Datacenter Compute Infrastructure (Starlink)
Builds and leads the team that sources GPUs, CPUs, storage, servers and racks for the AI training clusters and secures supplier allocation.
Palo Alto, CA or Austin, TX · on site · $155,000.00 – $215,000.00 (base salary) · posted 2026-07-24
Manager, Data Center Hardware & Infrastructure (TS/SCI)
Leads the team responsible for design, deployment, operations and lifecycle of data-center hardware across SpaceX sites.
Hawthorne, CA · remote policy not stated · $165,000.00 – $240,000.00 (base salary, plus a clearance differential once briefed) · posted 2026-08-21

What it likely means, and what to watch

InferenceBecause customers pay by the month and the Google agreement ties remedies to a delivery date, how fast delivered GPUs become fully operational now affects revenue directly, not just model training.
InferencePower, not chips, looks like the pacing item at Southaven: the temporary turbines must be removed by July 2027, so the permanent plant has to be commissioned on time or later chip waves would have less power than planned.

QuestionDid SpaceX deliver Google’s committed GPUs by September 30, 2026? Nothing read for this page confirmed it either way.
QuestionHow much of Colossus 2 is fully operational, rather than installed, by the end of 2026?
QuestionWill the permanent 1.2-gigawatt Southaven plant be running before the July 2027 deadline to remove the temporary turbines?

The wider industries that need this

The ten above are the loudest buyers, but the same job exists wherever scarce physical infrastructure has to become usable capacity on a date someone promised.

Hyperscale cloud
The biggest clouds are spending tens of billions of dollars a quarter on chips, buildings and power, and their own executives say customers still want more capacity than they can deliver. They need people who can say what is live today, what arrives next quarter, what is stuck, and who has to act. (Amazon, via Investing.com, 2026-07-30) (Microsoft Investor Relations, 2026-07-29)
Examples: Microsoft · Amazon (AWS) · Google · Oracle
AI labs
Labs sign multi-gigawatt deals years before the hardware exists, often with several suppliers at once, so contracted gigawatts and usable gigawatts are very different numbers. Someone has to track the whole portfolio and push decisions when a delivery slips, because training runs and launches are planned around those dates. (OpenAI, 2026-04-29) (Anthropic, 2026-04-20)
Examples: OpenAI · Anthropic · xAI
GPU clouds and neoclouds
GPU clouds report two power numbers: power under contract and power actually switched on. CoreWeave says revenue ramps once a cluster is delivered, so delays between contract and switch-on postpone revenue. (CoreWeave, 2026-09-17) (CoreWeave, 2026-08-11) (Nebius via SEC EDGAR, 2026-08-12)
Examples: CoreWeave · Nebius · Lambda · Nscale
Chipmakers
Chipmakers decide what AI hardware can exist at all, and their factories take years to build and ramp. A finished fab and a fab producing at volume can be most of a year apart, and customers plan clusters around those dates. Chip designers such as NVIDIA depend on those foundries. (NVIDIA, 2026-08-26) (TSMC, via Investing.com, 2026-10-15) (CNBC, 2026-10-15)
Examples: TSMC · Samsung · Intel · NVIDIA (designer)
Data-center developers and colocation
Developers build the buildings, power and cooling that AI hardware sits in, and report leased megawatts separately from ready-for-service megawatts. They need people who manage that gap building by building. (Applied Digital, 2026-10-07) (Crusoe, 2026-09-28)
Examples: Applied Digital · Crusoe · Digital Realty · Vantage
Power and utilities
Electricity is now a major limit on AI capacity, and utilities hold the queue. They are learning to separate firm requests from speculative ones and to estimate how long a site takes to ramp and how much of its contract it really uses. (International Energy Agency, 2026-04-16) (Dominion Energy, 2026-09-29) (Utility Dive, 2026-08-05)
Examples: Dominion Energy · AEP · PJM · ERCOT

06Why these roles can be remote or low-travel

Most of this work is coordination. The person in the role is reading forecasts from suppliers, dependency lists from construction and utilities, readiness status from commissioning crews, and asks from the teams waiting for capacity, then turning all of it into decisions for people across many sites, companies and time zones. That happens in dashboards, documents and meetings, not on a factory floor, which is why a good share of these postings are remote or hybrid.

The remote-tagged roles are desk and meeting work. Anthropic’s Reporting and Controls Lead keeps one milestone tracker current across every project and reconciles schedule data sent in by developers and contractors; OpenAI’s Development Planner turns compute forecasts and power schedules into a multi-year plan and builds dashboards. That work lives in schedules, spreadsheets and cross-site meetings, so it does not need to happen at a particular site, although Anthropic still expects at least a quarter of the time in an office.
Even some remote-tagged delivery roles ask for a lot of site time. Anthropic’s Capacity Deployment Lead is Remote-Friendly but expects 50%+ travel, and CoreWeave’s Commissioning/Quality Manager is remote-eligible only with 50% travel and residence near named regions. Microsoft’s GPU deployment program manager is remote with 50–75% travel.
Hands-on commissioning and site operations are on site because the duties are physical: xAI’s Mechanical Commissioning Lead directs functional tests on chillers, pumps and controls on a construction site, and CoreWeave’s Cheyenne Data Center Manager leads technicians through hardware repairs and joins on-call rotations.
Coordination does not automatically mean remote. NVIDIA’s and CoreWeave’s capacity program-manager postings list office cities and say nothing about remote work, and Amazon tags its Seattle and Herndon capacity postings as on-site. The dividing line is whether the job’s output is a forecast, a decision or a report, or a tested and energized system.

Be honest about the other half. Commissioning, installation and site operations are hands-on: someone has to be at the rack with the test equipment, and those roles are on site. The field has both, and most teams need both. Real examples of each, read on the dates shown:

Remote or remote-eligible postings

Anthropic — Reporting and Controls Lead, Data Center Capacity Delivery
Keeps one trustworthy milestone tracker current across every data-center delivery project, reconciles schedule data from developers and contractors, and builds delivery reviews and executive reporting.
Remote-Friendly, United States · Remote-Friendly; the posting also prints the company’s hybrid line of at least 25% office time · $320,000—$405,000 USD (annual salary) · posted 2026-08-24
OpenAI — Development Planner, Industrial Compute
Owns the multi-year plan for data-center shell capacity: turns compute forecasts, power schedules and cooling assumptions into site requirements, flags capacity gaps and stranded-capacity risks, and compares self-build, lease and colocation paths.
US – Remote · Remote (workplace type: Remote) · $225K – $385K, plus equity (salary as printed on the board) · posted 2026-09-15
Microsoft — GPU Deployment Execution Program Manager
Program-manages the execution of GPU deployments across Microsoft data centers.
Redmond, WA and four more; 0 days a week in office · remote, with 50–75% travel · USD $102,600 – $202,800 per year across the U.S.; USD $137,500 – $225,800 per year in the San Francisco Bay Area and New York City (base pay) · posted 2026-10-05
Anthropic — Capacity Deployment Lead – Data Center Operations
Confirms floor access, power, network and staffing are ready before racks arrive, then sets daily priorities and quality gates through delivery, power-on, burn-in and handover.
Remote-Friendly, United States · Remote-Friendly; 50%+ travel, on site at key points before and during delivery, deployment and close-out · $320,000—$405,000 USD (annual salary) · posted 2026-09-22
CoreWeave — Data Center Commissioning/Quality Manager
Reviews commissioning plans, witnesses functional tests of generators, UPS, electrical distribution and cooling on site, and reports progress and problems to leadership.
Must live near Northern Virginia, Dallas, Richmond VA, Livingston NJ or Atlanta · remote-eligible with 50% travel · $143,000 to $191,000 (base salary) · posted 2026-07-28

On-site postings

xAI — Mechanical Commissioning Lead – Memphis
Leads commissioning technicians through startup, functional testing and turnover of chillers, cooling towers, pumps and controls, and presents readiness reports to leadership.
Southaven, MS; Memphis, TN · full-time on site, with extensive travel to construction sites and extended hours during commissioning windows · pay not shown on posting · posted 2026-05-05, updated 2026-10-13
CoreWeave — Data Center Manager – Cheyenne, WY
On-site leader of a team of technicians doing hardware diagnostics, repairs and new installs, with regular inspections and audits to maximize uptime.
Cheyenne, WY · 100% on site; 24/7 environment with on-call rotations · $95,000 to $105,000 (base salary) · posted 2026-09-11
xAI — Program Manager, Data Center Construction – Memphis
Manages construction project managers delivering data-center builds from pre-construction through commissioning and handover, as owner’s representative to contractors and permitting authorities.
Southaven, MS; Memphis, TN · primarily on site in the Memphis area, with travel to other project sites · pay not shown on posting · posted 2026-07-01, updated 2026-10-13

07What people earn

Base salary is the fixed amount paid for the job; total compensation adds bonus and company stock, which at the companies above can be as large as the salary. Meta, for example, prints its range as base pay followed by “+ bonus + equity + benefits”, and CoreWeave says its total rewards add a discretionary bonus and equity awards on top of the base salary. The table shows ranges only where a source printed one, copied exactly, with the link and the date.

Role type	Example posting	Range as printed	Basis	Source
Capacity planner (compute demand)	Anthropic: Infrastructure Capacity Planner, Demand PlanningSan Francisco, CA \| New York City, NY	$320,000—$405,000 USD	annual salary (base)	company posting2026-08-11, updated 2026-08-21
Data-center capacity engineer (site level)	Meta: Data Center Capacity EngineerRayville, LA and six other locations	$111,010/year to $158,995/year + bonus + equity + benefits	base pay; the posting says the figures exclude bonus and equity	company postingundated page, read 2026-10-15
Technical program manager, data center / infrastructure	Anthropic: Technical Program Manager, Data Center InfrastructureSan Francisco, CA \| New York City, NY \| Seattle, WA	$365,000—$435,000 USD	annual salary (base)	company posting2025-11-26, updated 2026-10-07
Hardware (GPU) deployment program manager	Microsoft: GPU Deployment Execution Program ManagerRedmond, WA and four more; remote, 50–75% travel	USD $102,600 – $202,800 per year across the U.S.; USD $137,500 – $225,800 per year in the San Francisco Bay Area and New York City	base pay per year	company posting2026-10-05
Capacity operations program manager	Amazon (AWS): Senior Technical Infrastructure Program Manager, Infrastructure Capacity ManagementSeattle, WA	148,700.00 – 201,200.00 USD annually (Seattle)	base salary; sign-on and restricted stock units also mentioned	company posting2026-10-12
Capacity delivery / fleet program manager	CoreWeave: Senior Technical Program Manager – Fleet Delivery & Capacity ManagementNew York, NY / Sunnyvale, CA / Bellevue, WA	$182,000 to $242,000	base salary; discretionary bonus and equity also mentioned	company posting2026-09-16, updated 2026-10-12
Infrastructure program manager (senior / staff)	NVIDIA: Technical Program Manager, Infrastructure Capacity and DeploymentSanta Clara, CA	168,000 USD – 258,750 USD for Level 4, and 200,000 USD – 322,000 USD for Level 5	base salary per year, two levels; equity also mentioned	company posting2026-10-14
Data-center commissioning engineer / manager	CoreWeave: Data Center Commissioning/Quality ManagerNear Northern Virginia, Dallas, Richmond, Livingston or Atlanta	$143,000 to $191,000	base salary; discretionary bonus and equity also mentioned	company posting2026-07-28, updated 2026-10-12
Data-center operations / site manager	Amazon (AWS): Data Center ManagerNew Albany, OH	78,400.00 – 137,300.00 USD annually (New Albany, OH)	base salary; sign-on and restricted stock units also mentioned	company posting2026-10-14
Supply and demand planner for compute	Amazon (AWS): Senior S&OP Planner, Infrastructure Capacity & S&OP PlanningSeattle, WA	148,700.00 – 201,200.00 USD annually (Seattle)	base salary; sign-on and restricted stock units also mentioned	company posting2026-10-14

Ranges are copied exactly as each source printed them on the date shown. Company postings show base salary unless the posting says otherwise; pay-data sites mix base and total.

08Will this still matter in 10–20 years?

Honest answer: the chips will not last twenty years, but the problem will. Every generation of hardware so far has needed more power, more cooling and a faster network than the one before it, and each of those is a physical thing with a waiting list. The skill in this field is not knowing a particular GPU; it is turning scarce physical infrastructure into capacity that is usable on the day it was promised. That skill transfers to power plants, grids, chip fabs, transit systems and anything else that is expensive, slow to build and useless until the last piece is in place.

What credible forecasts say

FactUtility Dive reported on December 3, 2025 that BloombergNEF put U.S. data-center power demand at 106 gigawatts in 2035, 36% above its forecast earlier that year. (Utility Dive (reporting BloombergNEF), 2025-12-03)
FactPJM, the grid operator for much of the eastern United States, forecasts in its 2026 load report that summer peak load will rise by about 65.7 gigawatts over ten years and about 96.7 gigawatts over twenty, with data-center load cited as a driver in many of its zones. (PJM Interconnection, 2026-01-14)
FactGoldman Sachs Research forecast in February 2025 that global data-center power demand will rise as much as 165% by 2030 compared with 2023. (Goldman Sachs Research, 2025-02-04)
FactSEMI said on October 13, 2026 that global 300-millimetre chip-fab equipment spending will reach $149 billion in 2026 and $227 billion by 2029, with 20 new fabs or lines, excluding pure research fabs, starting construction worldwide in 2026 and 18 in 2027. (SEMI, 2026-10-13)
FactThe Charlotte Observer reported on February 10, 2026 that Duke Energy set a $103 billion five-year capital plan through 2030, with 4.5 gigawatts of data-center load secured in total, including 1.5 gigawatts under newly signed energy-services agreements. (Charlotte Observer via Yahoo Finance, 2026-02-10)
FactThe American Society of Civil Engineers’ 2025 Report Card graded U.S. public transit D and estimated that current funding from 2024 to 2033 would leave a $152 billion gap. (American Society of Civil Engineers, 2025)
FactThe U.S. EPA’s 7th Drinking Water Infrastructure Needs Survey found that drinking-water systems will need $625 billion over the next 20 years for pipe replacement, treatment-plant upgrades, storage and other key assets. (U.S. Environmental Protection Agency, page updated 2026-05-20)

The honest risks

Spending this large runs in cycles. Capacity can be built faster than the demand for it, hiring can freeze while buildings finish, and a company that bought too far ahead will slow down. Readers trust a page that says so, and the sources below do.

FactBain & Company estimated in September 2025 that about $2 trillion of annual revenue is needed by 2030 to fund the computing power expected to meet AI demand, and that the world is about $800 billion short even after AI-related savings. (Bain & Company, 2025-09-23)
FactERCOT, the Texas grid operator, told its board in September 2026 that after a state-ordered audit of data centers in its interconnection process it paused approvals to energize new large data-center loads of 75 megawatts or more until the audit is complete, with reports due by December 10, 2026. (ERCOT, 2026-09-14)
FactThe Oregon Capital Chronicle, republished on October 15, 2026, reported that Governor Tina Kotek ordered the state environmental agency to pause approving new air and water pollution permits for new data centers and expansions until July 1, 2027. (Oregon Capital Chronicle via The Daily Astorian and Yahoo News, 2026-10-15)
FactBloomberg reported on September 30, 2026 that federal regulators accepted but suspended PJM’s plan to procure new capacity for data-center growth until February 28, 2027, a pause of about five months. (Bloomberg via Energy Connects, 2026-09-30)
FactChicago Fed President Austan Goolsbee told Semafor in October 2026 that inflation tied to the AI boom is spreading into other parts of the economy, with prices for land, electricity, water, construction and electrical components used in data centers having surged. (Semafor, 2026-10-13 (updated))
FactBackground: TechCrunch reported on July 25, 2026, citing a Financial Times analysis, that U.S. tech companies had cut nearly 140,000 jobs since the start of 2026, even as several of the same companies raised their infrastructure spending. (TechCrunch (citing the Financial Times), 2026-07-25)

Where the skill transfers

Power generation. A turbine or reactor that has been ordered is not capacity until it is permitted, built, connected, commissioned and dispatchable. It is the same bought-versus-ready gap, and utilities are now building generation specifically for data-center load. (Charlotte Observer via Yahoo Finance, 2026-02-10) (PJM Interconnection, 2026-01-14)
Grid and transmission. Interconnection queues, load audits and capacity auctions are where paper demand becomes energized load. Someone has to verify which requests are real, sequence them and hold the schedule. (ERCOT, 2026-09-14) (Bloomberg via Energy Connects, 2026-09-30) (Lawrence Berkeley National Laboratory (U.S. Department of Energy), 2026-06)
Chip fabs. A fab is a multi-year build in which tools are installed, qualified and ramped before the site counts as capacity, and the industry is starting dozens of new ones. (SEMI, 2026-10-13)
Transit and rail. Rail systems move through install, test, commission and acceptance before they carry a single passenger: the same ready, available, used chain this page describes, on a different kind of hardware, with a national funding gap still to close. (American Society of Civil Engineers, 2025)
Water. Pipes, treatment plants and storage are long-lead, permitted, commissioned assets. They are slower than a GPU rack, but turning funded projects into working capacity is the same discipline. (U.S. Environmental Protection Agency, page updated 2026-05-20)

09The 8 Wastes of AI Capacity (preview)

Lean manufacturing names eight kinds of waste. Each has a twin in AI capacity, and naming them is the first step to seeing them in a readiness report. This is a framework in development; the full article is next on the list below.

01 · Lean: Waiting
Waiting
Racked GPUs waiting for power, cooling or a network they cannot use yet.
02 · Lean: Inventory
Inventory
Hardware in crates and warehouses, bought far ahead of the room that can hold it.
03 · Lean: Defects
Rework after failed acceptance
Re-cabling, re-flashing and re-testing after a burn-in or acceptance test fails.
04 · Lean: Overproduction
Idle accelerators
Accepted GPUs with no job to run, or reserved for a project that has not started.
05 · Lean: Motion
Poor allocation
Workloads moved between clusters, or run on hardware that is the wrong size for them.
06 · Lean: Transportation
Unnecessary handoffs
Readiness status passed through five teams before anyone can act on it.
07 · Lean: Over-processing
Over-processing
Reports, approvals and reviews that add steps without changing the decision.
08 · Lean: Non-utilised talent
Unused expertise
Commissioning crews waiting on paperwork; experts doing data entry.

Framework in development · October 2026

10The Control Tower (preview)

The Capacity & Readiness Control Tower is a demonstration dashboard. One screen that answers the five questions in the definition above: what exists, what is coming, what is blocked, when it will be usable, and which decisions are waiting on whom.

Funnel by site. How many GPUs sit at each stage, ordered to productive, for every site.
Blockers by cause. What is holding capacity back: power, cooling, network, parts, labour or paperwork.
Time to productive capacity. How long the last batches took from order to real work, and the trend.
Idle GPU-hours this week. Accepted capacity that ran nothing, and what that is worth at an assumed rate.
Forecast vs actual. The dates everyone planned against, next to the dates that happened.
Decisions waiting. Each open decision, its owner, its due date and what waiting costs.

Control Tower v1 · embed area

The demonstration dashboard will be embedded here when v1 is ready.

All demonstration data in the Control Tower is synthetic. It shows how the work would be run, not any company’s real numbers.

11Glossary

GPU: A graphics processing unit: the chip that does most of the arithmetic behind modern AI, because it can run thousands of small calculations at once.
Accelerator: Any chip built to speed up AI work, including GPUs and custom chips such as Google’s TPUs or Amazon’s Trainium.
Server (node): One computer in a data center, usually holding several GPUs plus the memory, network cards and cooling they need.
Rack: A tall metal cabinet that holds servers; a data center is rows of racks, each with its own power and cooling connection.
Cluster: Many servers wired together so they can work on one job at the same time, which is how large AI models are trained.
Data center: A building designed to run computers: it delivers power, removes heat, connects everything to the network and keeps it secure.
AI factory: A data center built mainly for AI work; the phrase treats compute as the product the building manufactures.
Hyperscaler: A company that runs data centers on a global scale and rents the capacity out, such as Amazon, Microsoft or Google.
Neocloud (GPU cloud): A newer company that rents GPU capacity by the hour or by contract, often built specifically for AI, such as CoreWeave.
Colocation: Renting space, power and cooling in someone else’s data center and putting your own hardware in it.
Capacity: How much computing a company can actually run, usually counted in GPUs, in GPU-hours or in megawatts.
Productive capacity: Capacity that is running real work most of the time at the performance that was paid for; the only kind that earns anything.
Idle capacity: Hardware that is powered on and accepted but running nothing.
Stranded capacity: Hardware that cannot be used because something around it is missing, such as power, cooling, network or a signature.
Headroom: Spare capacity kept on purpose so that growth or a failure does not stop the work.
Utilization: The share of time hardware is doing useful work; high utilization means little of it is idle.
GPU-hour: One GPU running for one hour; the common unit for pricing, planning and counting waste.
Training: Teaching an AI model by running it over huge amounts of data on a cluster, often for weeks.
Inference: Using a trained model to answer requests, which is what happens every time someone sends a prompt.
Commissioning: Powering up a new system and testing every part together, at full load, before anyone relies on it.
Burn-in: Running new hardware hard for a period to find the parts that fail early.
Acceptance: The formal moment the owner agrees that a delivery meets the contract, which starts payment, warranty and support.
Allocation: Assigning capacity to a team, a customer or a workload so it can be used.
Readiness: Whether everything a piece of capacity depends on is in place, so it can be used on the day it is promised.
Lead time: The wait between ordering something and receiving it; for transformers and switchgear it can be measured in years.
Interconnection queue: The utility’s waiting list for connecting a new large load, such as a data center, to the grid.
Megawatt (MW) and gigawatt (GW): Units of power; a gigawatt is a thousand megawatts, roughly the output of one large power plant.
Liquid cooling: Removing heat from chips with fluid instead of air, which the densest AI racks need.
PUE: Power usage effectiveness: total power into a data center divided by the power that reaches the computers; closer to 1 is better.
Capex: Capital expenditure: money spent on long-lived assets such as buildings, chips and power equipment.
TPM: Technical program manager: the person who coordinates a technical effort across many teams and keeps it moving toward a date.

12FAQ

BeginnerWhat is a GPU, and why does every company want them?: A GPU is a chip that can do thousands of small calculations at once. Modern AI is mostly that kind of arithmetic, so the companies building AI buy as many GPUs as they can get and race to put them to work.
BeginnerIf a company buys ten thousand GPUs, why can’t it use them the next day?: Because a GPU does nothing on its own. It needs a building, enough electricity, cooling, a fast network, software, testing and a team that knows it is there. Each of those comes from a different group with its own waiting list, and the slowest one sets the date.
BeginnerIs this the same as being a data-center technician?: No. Technicians install, cable and repair hardware on site. Capacity and readiness operations is the coordinating role: it tracks every site and supplier, knows what is blocked and why, and drives the decisions that unblock it. Both are needed; one is hands-on, the other is mostly planning and communication.
RecruiterWhat job titles should I search for?: Technical Program Manager with words like capacity, data center, infrastructure or deployment; Capacity Planner or Capacity Planning Manager; Hardware Deployment TPM; Infrastructure Program Manager; Commissioning Manager. The titles differ by company, so the pay section lists real examples with links.
RecruiterDo these roles need a computer-science degree?: Usually not. Postings ask for experience running large, multi-team infrastructure programs, comfort with data and schedules, and the ability to make a decision when the information is incomplete. People come from construction, utilities, semiconductors, transit, aerospace and manufacturing as well as from cloud companies.
RecruiterAre these roles remote?: Some are. Coordination roles that run forecasts, dependencies and readiness reviews across many sites are often remote or hybrid, because the work happens in dashboards and meetings. Commissioning and site-operations roles are on site. The remote section shows real postings of both kinds.
Hiring managerWhat does a strong candidate from outside AI bring?: The habits that make capacity productive are not AI-specific: readiness reviews that are decided before the date, dependencies tracked to an owner, acceptance defined before the test, and a single report leadership can act on. Someone who has carried a large physical program through commissioning has done this before, on different hardware.
Hiring managerWhat is the first thing a capacity and readiness lead should set up?: One honest picture of the funnel: how much capacity sits at each stage, per site, with the blocker and the owner for anything not moving. Most of the first wins come from making that picture visible, because the decisions it needs are usually cheap once someone can see them.
Hiring managerHow do you measure success in this role?: Time from order to productive capacity, and the share of accepted capacity that is actually running work. Both should fall and rise respectively, batch after batch, and both should be visible to the people paying for the hardware.
BeginnerWill this kind of work still exist in ten years?: The chips will change; the problem of turning scarce physical infrastructure into usable capacity will not. The outlook section gives the honest version, including the risks: spending cycles, overbuilding and hiring freezes.

13About this research

I’m Sina Nezhadian. I lead customer quality and root-cause problem solving on complex, safety-critical hardware programs, and I’m a Lean Six Sigma Black Belt.

At Alstom I worked on customer quality for BART’s Fleet of the Future, a ~$2.2B program for 775 railcars, and was Project Manager on the $4.9B LAX Automated People Mover, a driverless system, where I carried my scope through integration, testing, commissioning and operational-readiness handover.

I have not worked in a data center or on an AI infrastructure team, and this page does not claim otherwise. Productive Compute is independent research: I apply the readiness reviews, Lean thinking, risk management and data-driven delivery I used on transit programs to one question — how purchased AI hardware becomes productive capacity — and I say where the comparison holds and where it does not.

More about me →

14What’s coming next

01
The Productive Compute Brief Weekly
A weekly note on what moved in AI capacity: new sites, power deals, lead times, postings and the numbers behind them, each with its source.
02
The 8 Wastes of AI Capacity Article
The full article: each waste, how to see it in a readiness report, and the Lean countermeasure that fits.
03
Control Tower v1 Tool
The demonstration dashboard, embedded on this page, built on synthetic data.

Also in progress: fleet-scale quality

What a 775‑car fleet teaches about AI hardware reliability · in progress
8D for accelerator and server field failures: a working template · in progress
Customer non-conformances at fleet scale: from first report to verified closure · in progress
First-article inspection for AI hardware: what to check before volume · in progress
Test-readiness reviews: lessons from commissioning a driverless system · in progress
Lessons learned that travel: applying REX across sites and suppliers · in progress

Related research on this site

Fleet-scale quality: AI hardware reliability and customer qualityResearch hub · Customer quality
How the best tech companies find the real cause — and what it means for customer qualityFlagship study · Root cause · Customer quality
The requirement you find in testingRequirements · Test readiness · September 18, 2026
The constraint always movesConstraints · Continuous improvement · September 8, 2026

Sources (all accessed October 15, 2026)

Changelog

October 15, 2026First published. Before publication, relayed news reports and indirectly sourced leader names were removed, and a link to the fleet-scale quality hub was added.

Corrections and sources I missed: snezhadianpm@gmail.com.