Sixteen billion dollars in one week, paid by the company everyone else builds around.
Three and a half of it bought a seat inside the chips designed to replace it. Thirteen more bought the place open models already live.
The dominance holds.
Now it costs to keep.
I'm Ben Baldieri. Every week, I break down what's moving in GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
NVIDIA Spends $16 Billion in a Week as Buyers Rank Rivals Ahead of Blackwell
Buyers ranked the rivals ahead of its next GPU. It answered with two cheques and sixteen billion dollars.
NVIDIA put $3.5 billion into MediaTek's $3.9 billion convertible bond offering, deepening a partnership that opens NVLink to third-party custom silicon. With OpenAI, Google, Amazon, Meta and now Gulf buyers all designing their own accelerators, NVIDIA would rather sell the interconnect than lose the socket. AWS took the same trade days later, committing to NVLink Fusion alongside a further 2 million NVIDIA GPUs. Then NVIDIA confirmed it is buying Hugging Face for $12.93 billion, putting the platform 18 million developers use to find open models under the same roof as the chips. All of it landed as enterprise buyers were reported to rank non-NVIDIA accelerators 14 points ahead of Blackwell on their evaluation shortlists, and as AMD brought its Instinct systems online in Saudi Arabia on a 1GW build with Cisco and HUMAIN and won Europe's €387.8 million LUMI-AI supercomputer. None of it dents NVIDIA's quarter. All of it chips at the next one.
Why this matters:
The $3.5 billion buys NVIDIA a seat on chips it did not make. NVLink stays the default even on a rival's design, and AWS taking NVLink Fusion the same week shows the pitch works on the buyers who already build their own silicon.
Hugging Face is the front door to open models: 3 million of them, 500,000 datasets, 18 million developers, on $150 million of annualised revenue. For $12.93 billion NVIDIA now owns the shelf its competitors' models sit on, whatever it says about keeping the platform open.
Buyers ranking rivals 14 points ahead means the Blackwell premium is now negotiable. Every hyperscaler already shipping its own accelerator, from Meta's MTIA (#121) to Amazon's Trainium and Google's TPU, has a credible second source, and AMD's Saudi build and LUMI win put real deployments behind the optionality.
Lambda Lands $35 Billion From Anthropic, and NVIDIA Holds the Lease
The neocloud got the contract. Its chip supplier got the building.
Lambda has signed a reported $35 billion deal with Anthropic for 350MW at a Texas data centre, per the Wall Street Journal, with NVIDIA holding the lease on the building itself. The site is Hut 8's 525-acre Beacon Point campus in Nueces County near Corpus Christi, aiming for first power in Q1 2027 on the way to 1GW. Lambda goes into its IPO window with a contracted revenue line it did not have a month ago, and a landlord that also sells it the chips. For Anthropic it is the second neocloud offtake in a week, after the reported $45 billion, 460MW Nscale deal (#121), part of a run that since its confidential June IPO filing has taken in $19 billion with TeraWulf (#115), $9.1 billion with Riot (#119), and more than a dozen leasing letters of intent, on top of 10GW-plus of rented servers and a $200 billion commitment to Google. Neither company is public yet. They are already floating each other.
Why this matters:
The offtake is the float. Lambda was reported to be raising up to $3 billion ahead of its IPO (#121), and a $35 billion contracted line is what lets a banker price the book. It is the same move that carried Nscale towards its own listing last week.
NVIDIA holding the lease puts it on the hook for the building its own chips fill, the vendor-financing loop from its Nebius stake (#116) and its CoreWeave backstops. De-risk the neocloud, keep the next wave of demand pointed at NVIDIA silicon.
The whole neocloud IPO wave rests on one question: whether Anthropic's compute demand holds through 2029. Four landlords have already written their equity story around the answer.
Nscale Buys Into Its Own Customer to Win a $3.5 Billion Robot Contract
The neocloud is now a shareholder in the company it just sold the compute to.
Figure has committed $3.5 billion to Nscale for compute, with both sides saying they intend to scale past $6 billion, and Nscale took a strategic stake in the robotics firm on the way through. The deal covers up to 100,000 NVIDIA Vera Rubin GPUs from the second half of 2027 at Barstow, Texas, to train and serve Figure's Helix models, which founder Brett Adcock says improve the way every learned system does, with more data and more compute. It lands days after Anthropic's reported $45 billion, 460MW commitment to Nscale's West Virginia campus (#121), with a targeted $3 billion US IPO still ahead of it. Two contracts, two counterparties, one balance sheet that has to build for both.
Why this matters:
Taking equity in the customer is vendor financing pointed the other way. NVIDIA seeds the neoclouds that buy its chips (#116); Nscale is now backing the robotics firm whose cheques it is booking as revenue.
Physical AI is the demand story neoclouds have started underwriting. Humanoid robots need training compute on a frontier-lab scale, and 100,000 GPUs running from 2027 gives Nscale a book that does not ride solely on one frontier lab.
Nscale approaches its listing with two multi-billion contracts and a stake in one of the counterparties. That is a longer revenue line and a more tangled one, and public investors will price both.
GPT-6 Astra Is OpenAI's Largest Training Run, and Its Least Legible Model
The capability jump arrived with the reasoning traces switched off. OpenAI's chief scientist says the two go together.
OpenAI released GPT-6 Astra, the first model it has pretrained on more than 100,000 DBUs at Stargate and the first where earlier models did much of the supervision of the next one. It scores 72.6% on the OSWorld 2.0 offline subset against GPT-5.6 Sol's 65.7%, and gets there in roughly 40 minutes a task rather than 75. OpenAI also rates it the first model to reach Critical cybersecurity capability under its Preparedness Framework, after a perfect ExploitBench score and two zero-days found in testing, both company-stated. The capability came with a change to how the model thinks: opaque recurrence, which loops through reasoning internally instead of emitting a legible chain of thought. Redwood Research's Buck Shlegeris says pushing it further could totally destroy chain-of-thought monitorability. Chief scientist Jakub Pachocki's own framing is that capability and monitorability pull against each other, because a stronger model does more with fewer tokens. The Information reports that Anthropic and Google DeepMind are already discussing the technique.
Why this matters:
100,000 DBUs is the first hard figure attached to a frontier training run on Stargate. The capacity this whole build-out has been financing finally has a shipped model against it.
Earlier models supervising the next one shifts the mix towards raw cluster time and away from human-labelled data. That is a compute-demand story before it is a capability one.
No lab has shipped a frontier model with deliberately less legible reasoning before. If Anthropic and DeepMind follow, the industry gives up its main audit tool, and every buyer of inference inherits the consequence.
China's Domestic Chip Makers Hit the Commercial Turning Point
The Chinese alternative to NVIDIA stopped being a slide and started filing for IPO.
Four of China's domestic AI-chip makers are nearing commercial viability and moving into a differentiation phase, from prototypes to products meant to stand in for NVIDIA and AMD. The capital is following. Tencent-backed Enflame is targeting a $908 million Shanghai IPO, and SenseTime's chip spinout Sunrise reportedly raised RMB2 billion (about $280 million) at a RMB20 billion (about $2.8 billion) valuation. Huawei poured RMB121.4 billion into R&D (about $17 billion) in the first half alone, 25.9% of its revenue, most of it aimed at the compute stack export controls were meant to cut off. Even the memory is going domestic: CXMT started mass-producing LPDDR6 for Xiaomi's new foldable. Two years of export controls were meant to freeze this stack. They funded it instead.
Why this matters:
An Enflame listing gives a domestic accelerator maker a public balance sheet and a market price, plus a template for the YMTC and DapuStor IPOs already queued (#121). Public markets are how China scales this past prototype.
Huawei spending a quarter of its revenue on R&D is the state-backed reply to being cut off from NVIDIA. The compute, the memory (CXMT's LPDDR6) and the models that run on them, GLM-5.3 served entirely on Chinese chips (#121), are all being built inside the wall.
Every domestic part that reaches volume is a workload that never touches an NVIDIA GPU. The buyer optionality the West is only now testing already has a Chinese column, export controls or not.
Baidu Says Its GPU Clusters Pay Back in Three Years
The first hyperscaler to put a payback period on its AI compute did it in Beijing.
Baidu's CFO said AI-related revenue now tops half of company sales, with margins that could approach its search business as its GPU clusters pay for themselves in two to three years. It is the most concrete return figure any hyperscaler has put on AI capex, at a moment when Western investors still mark the spend as a question. Baidu has also split its cloud unit, moving model-as-a-service and its Qianfan platform into a separate infrastructure division and standing up a dedicated agent business. The sums are easier on its own Kunlun accelerators, which spare it the NVIDIA premium the West is still paying. Every Western hyperscaler has faced the same capex question for two years. Baidu is the first to answer it with a number.
Why this matters:
Alibaba got marked down for a $10.2 billion AI raise (#121). Baidu is telling investors the compute already earns its keep, and a two-to-three-year payback reframes the whole capex-fatigue trade.
Cheaper inputs move the bar. A payback clock running on Kunlun silicon beats one running on full NVIDIA pricing, which is why the figure came out of Beijing rather than Seattle.
The cloud reorg shows where Baidu thinks the margin is: agents and model-as-a-service, sold on top of infrastructure it now runs cheaply.
CPP and Equinix Buy atNorth for $4 Billion, and a Pension Fund Takes the Keys
The money buying the physical layer of AI now comes with a pension fund's name on it.
CPP Investments and Equinix completed a $4 billion acquisition of atNorth, one of the Nordics' largest data-centre operators, with the Canadian pension fund taking a controlling 51% for $1.3 billion, Equinix 34% for $895 million and Partners Group 10% for $260 million, behind a $4.1 billion financing package underwritten by European and Canadian lenders. atNorth brings eight operational sites across Iceland, Sweden, Finland, Denmark and Norway, with five more in development. For Equinix it is the first big platform buy of the AI build-out; for CPP it is infrastructure-grade exposure to compute demand without betting on any single tenant. Power and cooling are the two constraints everyone else is fighting over. This far north, they come cheap and green.
Why this matters:
Where neoclouds carry tens of billions in debt against depreciating GPUs (#119), CPP buys the freehold and the power. Patient pension money is a new kind of AI-infrastructure investor, betting on demand outlasting any one tenant.
Equinix taking 34% rather than the whole platform shows even the largest colocation operator would rather share the risk than carry it alone as capex climbs.
Every megawatt secured this far north is one not fought over in Virginia or Texas. The latitude does the cooling for nothing.
Everything Else
p.s. The GPU Daily now goes out every morning. Opt in below if you want the physical layer of AI in your inbox daily.