Two Million GPUs Is Not Two Million Instances

Yesterday I wrote about the top of the S-curve. This week’s capacity news is the other half of that question: what the suppliers are committing to build. Three headlines landed within days of each other. Oracle says it will raise $45 to $50 billion this year, in debt and equity, for capacity for Nvidia, Meta, OpenAI, AMD, TikTok and xAI.¹ AWS and Nvidia say they will deploy 2 million additional GPUs across 2027 and 2028.² Microsoft says it is tightening oversight of the neoclouds it already contracted.³ All three are talking about capacity. None are speaking the same units.

What a “GPU” means in a press release

Take the AWS number, because it is the one people repeat. It covers three GPU generations, Blackwell Ultra, Rubin and Rubin Ultra, across a two-year window, on top of a separate pledge of more than 1 million GPUs starting in 2026. As one analysis put it, the 2 million figure is “a count of accelerators across three generations, not as a Rubin count.”⁴ Neither company publishes what it counts as one GPU, and whether that means a die, a module, or a superchip changes the number materially.⁵

The packaging makes it concrete. A GB300 NVL72 rack, the current flagship, integrates 72 Blackwell Ultra GPUs with 36 Grace CPUs in one liquid-cooled rack wired into a single NVLink domain, plus 9 NVLink switch trays and power shelves.⁶ The CPUs are not add-ons; each GB300 superchip pairs one Grace CPU with two Blackwell Ultra GPUs, so roughly a million CPUs arrive with any 2-million-GPU order of that generation.

Run the rack arithmetic and the unit fog gets easier to see. Two million accelerators divided by 72 per rack is about 27,800 racks, but the span includes denser Rubin Ultra racks at roughly twice that density, so the same headline number is between about 14,000 and 28,000 racks. That is a plan, not a rate card. What lands in a customer’s account is instances, and the gap between paid-for accelerators and schedulable, usable capacity is the gap the announcement does not describe. A rack that arrives at a site that cannot power or cool it is not delivered.⁷

Neither company publishes what it counts as one GPU, and the difference between a die and a package is millions of racks of arithmetic.

Oracle is raising the money the stock won’t give it

Oracle’s announcement is the financing half of the same story. In a statement on February 1 it said it would raise $45 to $50 billion of gross proceeds during calendar 2026, roughly half in equity-linked instruments, the rest in senior unsecured bonds.⁸ That was before its fiscal first quarter showed remaining performance obligations of $664 billion, up $209 billion year over year, alongside free cash flow of negative $5.4 billion on $28.5 billion of capital expenditure.⁹¹⁰ Cloud infrastructure revenue grew 121% that quarter.

Meanwhile the market is repricing the risk. Oracle closed at $116.10 on November 6, against a 52-week high of $322.54, about 64% below the peak.¹¹ It is funding the largest buildout in its history in the capital markets while its own equity reprices whether the contracts behind that buildout will pay. Whatever you think of the bet, forging ahead is accurate. Amazon is running a quieter version of the same play: its third-quarter filing disclosed a supplier loan facility of up to about $80 billion, secured against the hardware it buys.¹²

Microsoft is checking the receipts

Microsoft’s move reads differently, and I think it is the smarter one of the three. Bloomberg reported on November 5 that Microsoft is tightening oversight of projects run by Nebius, IREN and Nscale, sending its own people to construction sites, checking the work against its specifications, and debating whether to sign new neocloud deals at all.¹³ The pullbacks are concrete: it walked away from a 1.35 gigawatt letter of intent at Nscale’s West Virginia campus and cancelled a roughly $3 billion UK deal over power problems. The partners’ disclosed contracts total about $70.9 billion per their own SEC filings, and that is the scale now under review.¹⁴

The delays back the review up. IREN said deliveries for Microsoft slipped after Microsoft requested network-design changes, pushing a revenue milestone into early 2027.¹⁵ Amy Hood, Microsoft’s finance chief, trimmed the calendar-2026 capex forecast to about $170 billion, citing third-party capacity arriving late.¹⁶

My read: the hard part of GPU capacity was never signing the contract. It is power, cabling, construction, and someone on site with a checklist. Oracle and Amazon are committing faster. Microsoft is verifying slower. In a market where every announced number is a plan rather than an instance, slower verification is the more intelligent approach, and the one that shows up in delivered capacity first.

Sources