AI Data Center Buildout: Nvidia Opens Vera Rubin Rack Designs

The rack is now the bottleneck

On October 12, Nvidia went to the Open Compute Project Global Summit in San Jose and gave away the blueprints for the hardware it plans to sell the most of next year. The company contributed its Vera Rubin rack and compute-tray designs to the OCP through an updated MGX architecture. The goal, in Nvidia's words, is a common foundation for companies building next-generation "AI factories."29 The new MGX 1.2 specification supports racks drawing up to 250 kilowatts. It adds blind-mate liquid cooling, a liquid-cooled 5,000-amp busbar and fanless, liquid-cooled compute trays.2330

One line explains the move. Joe DeLaere, who runs Nvidia's AI infrastructure products, said the limit on AI factory buildouts is how fast the industry can manufacture and deploy them.26 With published drawings, a maker of cooling plates and a supplier of power rails can build to the same specifications at the same time. The chips inside the rack still belong to Nvidia.26 Nvidia made the same kind of contribution for its Blackwell rack two years ago.26

The OCP package covered more than the rack. Nvidia released a Vera CPU rack specification, open-sourced four DSX software components, showed Spectrum-6 switches at 102.4 Tb/s, introduced 400Gb/s-per-lane USFP optics, and published facility standards with hyperscalers.30 One of those standards requires a data center to return to near-full load within two seconds of a grid disturbance.30 Meta said it will use Spectrum-6 Ethernet switches to build larger Vera Rubin clusters on open networking. Nvidia says its Spectrum-X Multiplane technology can scale to AI factories of more than 100,000 GPUs.23

The main takeaway is that Nvidia is no longer just selling chips into this buildout. It now treats the building, the power and the supplier network as part of what it sells.

What Vera Rubin is, and when it arrives

Rubin is a full-system redesign, not a chip refresh. Nvidia describes Vera Rubin as a rack-scale platform built from several co-designed chips. The NVL72 rack holds 72 Rubin GPUs and 36 Vera CPUs and is rated at 3,600 petaflops of NVFP4 inference. Nvidia claims that is five times Blackwell's throughput at a tenth of the cost per token.21 Each Rubin GPU uses a dual-die TSMC 3nm design with 336 billion transistors and 288GB of HBM4 running at 22 TB/s.27 A companion LPX rack, based on Groq 3 LPU technology, handles low-latency decoding. Nvidia claims it delivers 35 times more throughput per megawatt than Blackwell on trillion-parameter models.21

Coverage of the timeline is less consistent. One account says Rubin entered full production at CES in January.21 Another dates the formal "ramping into full production" announcement to May 31, with broad customer shipments starting this fall.24 Forbes places Jensen Huang's full-production announcement in June.26 At least one outlet still reports Rubin GPUs sampling in Q4 2026, with volume production in Q1 2027.22

The most consistent reading is this: early NVL72 systems began reaching partners such as CoreWeave over the summer, and wider shipments are happening now.24 Broad access through neocloud providers looks more like a 2027 event.21 That gap explains why Nvidia is focused on deployment speed. The chips exist. The racks, cooling and buildings around them are what decide when they start earning money.26

The competition has also moved to the rack. AMD's Helios design pairs 72 MI455X GPUs, each with 432GB of HBM4, against Nvidia's 288GB per GPU. As of mid-2026, Helios was still at the reference-design stage.24 Forbes notes that Meta both designed AMD's rack and is doing its open networking work on Nvidia switches. In effect, the two vendors now share the building and compete for the rack inside it.26

The capex numbers: big, rising, and counted differently

Everyone agrees the hyperscaler spending is enormous. The disagreement is over the exact figure, and it comes down to who is counted and what is included.

For the four largest spenders (Amazon, Microsoft, Alphabet and Meta), Forbes puts 2026 capital spending at $700 billion to $725 billion, up about 75% from $410 billion last year.1 Tracking the companies' latest guidance gives Amazon about $220 billion, Alphabet $195 billion to $205 billion, Meta $130 billion to $145 billion, and Microsoft about $175 billion after lease accounting changes. The midpoints add up to about $733 billion.410 On Microsoft's earlier accounting basis, the four-company total would be closer to $747.5 billion.4 Adding Oracle pushes the totals higher still. Morgan Stanley raised its five-company 2026 projection to $805 billion7, and Bank of America puts the figure near $860 billion.9 UBS uses a broader definition and puts total hyperscaler capex at $1.009 trillion this year.6

The forecasts for next year all point up. Goldman Sachs expects AI capex at the five largest US hyperscalers to rise more than 50% to $1.2 trillion.2 S&P Global Ratings expects combined hyperscaler spending to exceed $1.3 trillion by 2027.6

Not all of that growth is new capacity. Dell'Oro Group found worldwide data center capex rose 92% year over year in Q2, while revenue from the chips and components inside those facilities jumped 182%. Higher memory and storage prices lifted server prices.6 Microsoft's CFO said $25 billion of the company's roughly $190 billion 2026 plan reflected higher component prices.4 Amazon's increase to $220 billion was also tied to memory costs.4 Some of the eye-catching growth is inflation in the bill of materials, so the headline numbers overstate how much new compute is actually being built.

Who pays, and with what

The financing has changed, and the coverage is notably consistent on this point. Capex is expected to use up nearly all of the big hyperscalers' operating cash flow this year. That turns companies once known for light, cash-rich business models into capital-intensive infrastructure operators.1 Bank of America projects aggregate free cash flow falling from about +$180 billion in 2025 to about -$64 billion in 2026 and -$144 billion in 2027.9 S&P expects all six of the largest hyperscalers to have negative free operating cash flow in both years.6

Debt is filling the gap. Morgan Stanley estimates hyperscalers will borrow about $400 billion in 2026, double the 2025 figure.7 Bank of America counted $344 billion in AI-related bond issuance by early August.18 The Wall Street Journal reported about $3 trillion in off-balance-sheet purchase and lease commitments across nine AI spenders, disclosed in footnotes rather than on the balance sheet.18 Nvidia itself sits inside a web of circular deals. Its stakes in OpenAI, CoreWeave, Intel and others come with commitments to buy Nvidia hardware.18

The demand case is real, but it is concentrated. CFRA counted more than $2.5 trillion in combined cloud backlog in Q2, up from about $750 billion three quarters earlier. A meaningful share of that is tied to a few large AI customers.19

Power, labor and the physical ceiling

Over the longer term, the estimates are huge and do not line up. Bain projects $5 trillion to $6.5 trillion of spending to add about 150 gigawatts by 2030, nearly tripling global capacity. Bain also warns that power, chips, labor and permits are all constrained at the same time.13 McKinsey's base case is about $5.2 trillion in AI-related data center investment through 2030, with roughly $3.1 trillion going to chips and computing hardware.14 A Brookings study by Columbia's Stijn Van Nieuwerburgh projects $10.3 trillion of US AI infrastructure spending from 2025 to 2032, about 3.6% of GDP per year.15

Power is the limit that money alone cannot fix. The CEO of data center builder DataBank says no power is available for 2026 through 2028 because it has all been reserved.12 The IEA expects data center electricity use to roughly double from about 485 TWh in 2025 to around 950 TWh by 2030.14 Meanwhile, Hyperion, Meta's five-gigawatt Louisiana campus, has local residents worried about who will pay for new gas plants if the tenant leaves.11

The read

The OCP announcement captures the state of the market. Chip demand is not the problem. Nvidia is shipping its most ambitious platform yet into a market whose spending keeps beating forecasts. The risk has moved to everything around the chips: grid connections, 250-kilowatt liquid-cooled racks, memory prices, and a financing structure that increasingly depends on borrowed money. Opening the rack designs is Nvidia's attempt to shorten the time between a chip leaving the fab and running in a data center. That makes sense for Nvidia, because every rack built to its drawings is filled with its silicon. It does nothing to reduce the bigger exposure the buildout is creating: an industry spending more than it earns on the expectation that AI revenue will eventually catch up. By most accounts, 2027 is the year that expectation gets tested.9