In partnership with
For two years the shape of the frontier AI model market barely moved:
The best models came from a handful of Western labs, they stayed closed, and reaching the frontier meant paying for a login. Open models were the consolation prize, free to run, several steps behind, and priced low because they had to be.
This week, the competition broke that arrangement.
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.
Today's issue is brought to you by Rafay
The GPU runs on sponsorships. If you value independent, no-fluff analysis of the AI infrastructure market, the sponsors are who make it possible. I'd appreciate you checking them out.
The World's Leading Neoclouds Run on Rafay
Rafay helps neoclouds, sovereign AI clouds, and enterprises turn GPU infrastructure into self-service, governed AI cloud services, from Token Factory and inferencing to Kubernetes, SLURM, bare metal, and VMs.
Moonshot's Kimi K3 Lays Claim to the Frontier in the Open
A Chinese lab says its next open model already matches the labs charging a premium for access.
Moonshot AI announced Kimi K3 on 16 July: a 2.8-trillion-parameter multimodal model with a one-million-token context window and open weights landing on 27 July, billed as the largest open-weight model yet from China. The benchmarks are Moonshot's own for now, so read them with that in mind. On Arena it topped the front-end coding board, ahead of GPT-5.6 Sol, and edged the standard version of Anthropic's Opus 4.8; on GDPval-AA v2 it scored 1,668, third behind Claude Fable 5 Max and GPT-5.6 Sol Max. The number that matters more sits on the pricing page. At $3 input and $15 output per million tokens, Kimi K3 costs what a mid-tier American model costs, not the usual Chinese discount. Moonshot wants the benchmark crown and the margin.
Why this matters:
Frontier parity is no longer a Western, closed franchise, and an open model that self-hosts erodes the API lock-in OpenAI and Anthropic depend on.
It’s widely known that both Anthropic and OpenAI heavily subsidise their users. With benchmark numbers like these at Sonnet-level pricing, those subsidies will likely no longer be enough.
If adoption rates and revenues suffer a meaningful hit (especially pre-IPO), expect some heavy national security-adjacent messaging. Because nothing inspires product confidence quite like arguing for less competition.
TeraWulf Is Raising $3.5B to Build Anthropic a $19B Home
A former bitcoin miner is borrowing billions to build a home for a lab that could fund the building itself.
TeraWulf spent this week under fresh scrutiny over its roughly $19 billion, 20-year Anthropic lease, which Data Center Frontier says puts the former miner's brownfield strategy to the test now that it is landlord to one of the largest compute tenants in the market. To build it, TeraWulf is raising $3.5 billion in debt, with Morgan Stanley expected to lead. It's the pivot playing out across the sector: crypto sites with power already in hand, re-papered as AI hosting, with Soluna filing an 8-K this week doing a version of the same thing. The structure is the point. Anthropic keeps the capex off its own balance sheet and pushes the debt, the construction risk and the power procurement onto a counterparty that in effect takes the first loss if the lease ever wobbles. It's the second month running we've watched an Anthropic build-out financed on someone else's books.
Why this matters:
Frontier labs have decided owning buildings is someone else's job. The lab signs the lease, the miner raises the debt, the bank underwrites the tenant.
The risk concentrates in the counterparty. A $19B lease is only as good as Anthropic's revenue in 2029, and TeraWulf carries the construction and refinancing risk until then.
Reflection Commits $1B to Nebius for Nvidia's Latest Chips
An open-weight lab just handed a neocloud a billion dollars rather than wait years to build its own.
Reflection AI signed a $1 billion capacity deal with Nebius for access to Nvidia's newest GPUs. Reflection, founded in 2024 by two ex-DeepMind researchers, carries an $8 billion valuation and $2.6 billion raised from backers including Nvidia and Sequoia. We last tracked it in Issue #112. A billion-dollar cheque from a two-year-old lab is now an ordinary way to secure frontier capacity, and the independents are the ones cashing it: Nebius here, CoreWeave and Crusoe before it.
Why this matters:
Neocloud demand is set by a handful of labs writing enormous capacity contracts. Nebius just booked one that reprices its backlog.
It is the neocloud pattern from Issue #112 holding: the independents, not the hyperscalers, keep winning the frontier-lab compute.
An open-weight lab renting rather than owning tells you where the margin sits: in capacity, not the model (see Kimi K3 above).
Microsoft Coaches Its Sellers to Talk Down OpenAI and Anthropic
The largest backer of OpenAI is reportedly teaching its sellers to argue against it.
Microsoft executives, per TechCrunch, told salespeople to compare OpenAI and Anthropic unfavourably against its own models and pitch "the full end-to-end system." EVP Jacob Andreou reportedly put Copilot against Claude, calling Claude "slower and less accurate" inside Microsoft's apps. Days earlier, OpenAI had named GPT-5.6 the "preferred model" for Microsoft 365 Copilot, even as Bloomberg reported Microsoft swapping in its in-house MAI models to cut costs. Read it as an infrastructure move: every query routed to MAI is inference Microsoft runs on its own silicon, against the $31.9 billion in capex it spent last quarter.
Why this matters:
The model layer is commoditising in real time, and whoever owns the distribution owns the customer. Kimi K3 in the open only speeds that up.
Set it against the frontier-race question from Issue #113: if MAI can stand in for GPT-5.6 without users noticing, OpenAI's best channel becomes its least defensible.
Follow the demand, not the drama. Repatriated inference means more Microsoft-owned capacity and less third-party spend.
Japan Lines Up a 27,500-Rubin National AI Factory with Nvidia
The next-generation Nvidia silicon is getting a government-backed home, and it isn't in America.
Nvidia, the Japanese government and a Noetra-led consortium launched what they call the world's first national AI infrastructure on 16 July, an AI factory to be built on 27,500 Rubin GPUs and 13,750 Vera CPUs. That would make it among the earliest Vera Rubin deployments at scale anywhere, and a sovereign one: state-backed, aimed at Japan's physical-AI and robotics workloads, not rented on the open market. Rubin racks reach volume production only in the second half of this year, so this is a commitment, not a running cluster.
Why this matters:
One of Rubin's first flagship commitments going to a sovereign programme rather than a US hyperscaler shows where Nvidia's next generation gets allocated: to governments willing to underwrite it.
Sovereign AI is a buyer class in its own right now. Japan is buying compute the way it once bought reactors, and the export-control thread from Issue #111 decides who else gets to.
The Gulf and Japan are now competing to be the first non-US homes for frontier silicon. Watch which government secures Rubin capacity next (see the UAE export easing below).
Google's 2.7GW "Project Tembo" Surfaces in Wyoming
A single campus is about to draw more power than some countries.
Google is the customer behind a 2.7GW data centre campus near Cheyenne, Wyoming, now renamed Project Tembo. It would be one of the largest single loads on the US grid. The same week, Google also committed to the Steel River Energy Center in Arkansas as anchor investor and offtaker, taking 100% of its initial output, a project scaling to 2.5GW of solar and 2.9GWh of storage. The gigawatt has replaced the megawatt as the unit of ambition, and securing the electrons is now the hard part.
Why this matters:
Power procurement is the real hyperscaler moat. Whoever locks in firm gigawatts first gets to build first.
The capex is relentless: Amazon alone spent $44.2 billion last quarter, and campuses like Tembo are where that money lands.
A 2.7GW load in one county is exactly the strain that pushed New York to a moratorium this week (below).
Blackstone, Apollo and KKR Put $5.34B Into Williams' Gas Plants
The biggest names in private capital spent the week backing turbines, not chips.
Williams secured $5.34 billion from Blackstone, Apollo and KKR for five behind-the-meter power projects serving data centres. The investors take 49% equity; Williams keeps 51% and control. The commitment splits into $4.4 billion of growth capex and $0.9 billion of additional consideration, against 2.6GW of announced capacity and a backlog north of 6GW. Blackstone last showed up in the power coverage in Issue #110. The real money now sits one layer below the GPU, in the generation that decides whether the GPU ever switches on.
Why this matters:
The scarce asset is a megawatt you can actually connect, and infrastructure capital knows it. Behind-the-meter gas skips a grid queue measured in years.
Same speed-to-power logic behind the Crusoe builds and Musk's turbine buy. Generation is a competitive weapon now, not a utility bill.
The clean-energy tension is unresolved. New gas for AI collides with every net-zero pledge, and this week's solar deals don't close the gap.
Everything Else
p.s. I launched a new daily update from The GPU this week. Opt-in below if you want me to send you more emails!