The OpenAI-Broadcom deal is a milestone map, not a data center

On October 13, 2025, OpenAI and Broadcom announced a multiyear collaboration to deploy ten gigawatts of OpenAI-designed AI accelerators. The joint announcement targeted the second half of 2026 for the start of rack deployment and the end of 2029 for completion.

The number describes an industrial ambition, not a data center already built. The text publishes no price, chip count, locations, wafer manufacturer, annual energy, or performance. Reading it rigorously requires separating five layers: existing agreements, term sheet, design, physical deployment, and operation. Each produces different evidence.

The announcement contains two instruments, not one contract

OpenAI and Broadcom said they already had agreements covering co-development and supply of the accelerators. They added that they had signed a term sheet for deploying racks incorporating those chips and Broadcom networking. The announcement publishes none of these documents or their economic obligations, exit conditions, or minimums.

A term sheet records a basis for negotiating or structuring a transaction; its force depends on its clauses and governing contract. It should not automatically be described as a final purchase order or a mere conversation. Here, the supportable statement is narrower: development and supply agreements existed, while rack deployment had a term sheet.

An audit sheet should assign a status to each promise: announced, agreed, contracted, financed, manufactured, installed, energized, and operational. Jumping from “collaboration” to “available capacity” erases years of work and conditions. The next useful document would be one that turns the schedule into orders, volumes, payments, and acceptance criteria.

Designing an accelerator does not mean manufacturing everything

OpenAI will design the accelerators and systems. Broadcom will collaborate on development and deployment, provide Ethernet for scale-up and scale-out, and supply PCIe and optical connectivity. The racks were intended for OpenAI facilities and partner data centers.

The announcement does not identify the chip manufacturer, process node, unit count per gigawatt, or the builder and operator of each facility. “Custom chip” here means a design directed by OpenAI; it does not show that the company owns the fab, grid connection, or all supporting software.

A custom accelerator can tune memory, number formats, data movement, and communication to particular workloads. Benefits do not appear merely because it is custom. They must be tested through sustained performance, model quality, availability, energy, and total cost against an alternative doing the same work.

Ten gigawatts is neither a bill nor annual energy

The US Energy Information Administration explains that a watt measures power at an instant, while a watt-hour measures energy used over time. One gigawatt is one billion watts. Without operating hours and utilization, power cannot be converted into annual energy consumption.

The announcement refers to ten gigawatts of accelerators but does not define the measurement boundary. It does not say whether the figure is nameplate chip power, complete rack power, or facility capacity. Nor does it provide the additional demand from host processors, networking, storage, cooling, and electrical conversion. Presenting ten gigawatts as total consumption would add a definition the source lacks.

The unit does not reveal chip count either. That depends on power per accelerator, rack design, and how the chip evolves over four years. Two systems drawing the same power may deliver different memory, communication, utilization, and accepted work.

Cost is another axis. The parties published no contract value. Multiplying gigawatts by a general estimate creates a hypothetical number, not this agreement’s price. A cost calculation would need included equipment, construction, electricity, financing, maintenance, duration, discounts, and allocation between parties.

The network is part of the computer

Large models distribute data and computation across many accelerators. Scale-up connects devices collaborating tightly within a domain; scale-out joins more systems to expand the cluster. Broadcom said both levels would use Ethernet and that the racks would include its connectivity portfolio.

“Ethernet” alone does not measure performance. Useful bandwidth, latency, loss, congestion, topology, and collective software behavior matter. A fast accelerator may wait for the network; an oversized network may sit idle when memory or computation is the limit. The system must be evaluated as a whole.

The relevant test is not an isolated maximum. Training should record time to a model meeting a quality threshold, failures, restarts, and utilization. Inference should measure accepted tokens per second, percentile latency, energy, and cost per request. Figures must include racks, networking, and hosts, not only silicon.

The schedule is a sequence of tests

“Start in the second half of 2026” and “complete by the end of 2029” are future targets published by the companies. A verifiable schedule decomposes them into design closure, first fabrication, packaging, working sample, rack validation, power availability, installation, acceptance, and commercial operation.

Each milestone addresses a different risk. A chip can work and arrive late; a rack can be installed without power; an energized facility can miss utilization targets; a system can run models without improving cost. Reporting that a project is “under way” without naming the milestone obscures what has been achieved.

Accumulated capacity also needs a date. Ten gigawatts at the end of the period does not mean ten throughout every year. A correct chart shows operational power added by quarter, contracted capacity awaiting delivery, and retired systems. This avoids adding maximums from separate announcements as if they were simultaneous.

Reported demand does not prove execution

OpenAI linked the collaboration to demand growth and said it had passed 800 million weekly active users. This is a company-reported figure that does not show usage per person, product workload, or the split between training and inference. It is attributed context, not reproducible capacity planning.

Justifying ten gigawatts would require a workload forecast: requests, tokens, models, service level, expected efficiency, and reserve margin. That information may be confidential, but its absence limits inference. The announced size signals the parties’ plan; it does not prove that the amount is necessary, sufficient, or financed.

A matrix turns the promise into follow-up

The columns are date, instrument, obligation, owner, scale, condition, evidence, and status. The first row covers development and supply agreements; the second covers the rack term sheet; later rows cover fabrication, network, power, installation, and operation. Cost, units, and locations absent from the announcement remain marked “unknown.”

When an update arrives, it does not replace the headline; it fills the matrix with documents and measurements. An installed rack proves construction; acceptance proves specifications; live workload proves use; cost and performance prove utility. None substitutes for the others.

The OpenAI-Broadcom collaboration matters because it tries to coordinate silicon, networking, and deployment at enormous scale. A rigorous reading does not need an invented bill. The transferable skill is following the chain from paper to useful work: what was signed, who must do it, which unit was measured, which milestone was reached, and what result it produced.

This discipline also prevents a misleading total. If OpenAI announces capacity with several partners, maximums should not be added until it is clear whether they are incremental, alternative, overlapping, or intended for different periods and workloads. Each agreement must retain its unit, date, condition, and measurement boundary. Only observed operational capacity can be consolidated, and reserved capacity must still be separated from utilized capacity. A supplier portfolio shows diversification; it does not prove that every machine will run together or that power and demand exist to keep them busy.