In partnership with
An unreleased AI model just did original mathematics.
At least, that is the claim. The same lab overreached on exactly this last year and quietly walked it back. But, this time will surely be different.
Right?
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 stays independent because sponsors cover the bills, not because they shape the copy. Rafay works the orchestration and GPU-management layer beneath all of this, so if you build, buy or fund the physical stack of AI, they are worth a look.
AI Infrastructure Leadership Summit
Hosted by Rafay, join fellow tech executives for this two-day summit bringing together leaders from the world's most innovative neoclouds, telcos, NVIDIA Cloud Partners, sovereign AI initiatives, and enterprise AI platforms to share real-world strategies for commercializing AI infrastructure, expanding margins, and building the next generation of AI businesses. *Registrations subject to approval
OpenAI's Unreleased Astra Model Cracked Ten Open Maths Problems
OpenAI says its next model cracked ten open problems in mathematics, and it published proofs a machine can check.
OpenAI published results on 1 August from an internal version of Astra, its next major model, that resolved or advanced ten long-standing problems in mathematics and theoretical computer science: a disproof of Connes's rigidity conjecture, new bounds on high-dimensional sphere packing, fresh lower bounds on multicolour Ramsey numbers. For each, the model wrote a formal proof in Lean, the assistant mathematicians use to check their own work, so the results verify by machine rather than on trust. Astra is not released. OpenAI says the tokens to find all ten cost about $2,000.
Why this matters:
A proof a computer can verify is cheap to trust, once a human confirms the problem was worth solving. Make that loop work and AI-for-science turns into a standing inference bill.
Now the part worth scrutiny: a Lean proof shows the logic compiles, but says nothing about whether the problem was really open, or whether the answer matters, and no mathematician has signed these off.
We have also seen this film flop before: last October an OpenAI researcher took a victory lap for GPT-5 "solving" open Erdős problems already sitting in the literature, and the claim was deleted within days. Until the field checks the work, "ten open problems" is a marketing line with a compute bill attached.
OLIX Raises $312M at a $3.3B Valuation, Backed by Arm
Arm just backed a two-year-old London startup trying to beat Nvidia at inference with light instead of copper.
OLIX, founded in 2024 by 25-year-old James Dacombe, has raised a $312M Series B at a $3.3B valuation, triple its worth six months ago. Fundomo led; Arm, Hudson River Trading and Netflix co-founder Reed Hastings joined, and the UK government's Sovereign AI fund increased its stake. The bet is optical: replace the copper between inference chips with light, and drop HBM from the design. First silicon is due in 2027.
Why this matters:
Tripling to $3.3B in six months, for a company with nothing shipping, shows how much capital will chase a credible alternative to Nvidia's inference economics before a single chip exists.
The UK Sovereign AI fund doubling down puts state money behind a British silicon champion. Arm also joining says the company inside an increasing number of inference servers wants the fabric between them too.
Dropping HBM is an aggressive strategy. If optical interconnect works at scale, OLIX sidesteps the memory bottleneck that HBF and CXL are only trying to widen.
Anthropic Commits $10 Billion to Volta, a Neocloud Founded This Year
Anthropic just handed $10 billion to a cloud company that did not exist a year ago.
Anthropic signed a $10 billion, six-year deal to buy compute from Volta, an AI-cloud startup that came out of stealth this week backed by Nvidia and Dell at a reported $2.4 billion valuation. The capacity is 133MW at Bitdeer's Tydal campus in Norway, part of a $4.7 billion, 16-year lease Volta signed with the crypto-miner, running Nvidia's Vera Rubin systems and reportedly underwritten by JPMorgan. It is Anthropic's third compute grab in months, after SpaceX and Amazon, and a ten-figure cheque to a company that barely exists.
Why this matters:
When a lab pre-commits $10 billion to a one-year-old startup, the binding constraint is who underwrites the buildout, and JPMorgan behind it is Nvidia's ecosystem borrowing against tokens not yet sold.
Crypto-miners are cementing their position as the landlords of AI. Bitdeer owns the site, the power and the grid connection, the same pivot turning idle mining capacity into GPU halls across the Nordics.
Anthropic is buying compute anywhere it can, after its SpaceX and Amazon deals and the $40bn-scale alignment in Issue #116. It hedges every supplier because the one thing it cannot do is run short.
Alibaba Ships a 2.4T Qwen as DeepSeek Undercuts the Field
The same week OpenAI showed what the frontier can do, China set out to make it cheap.
Two releases landed within days. Alibaba launched Qwen3.8-Max, a 2.4-trillion-parameter flagship it says matches the top Western models on agentic tasks, and its shares rose on the news. DeepSeek followed with V4-Flash, rated by independent benchmarks the cheapest capable model to run, around 100 times less per token than Anthropic's Claude, served through China's National Supercomputing Internet. One went big, the other went cheap, and both point one way: the gap with the US labs keeps narrowing while the price of a frontier-class model keeps falling.
Why this matters:
Cheap open models pull inference down-market. When a capable model costs a hundredth of the incumbent, the world runs far more tokens, and most land on whatever hardware is cheapest to serve.
Qwen at 2.4 trillion parameters is the capability signal: a Chinese lab shipping at that scale, rewarded by the market, means the frontier is no longer a two-lab race.
Astra and this are mirror images: one lab spends to push capability out, a national ecosystem drives the price down, and the compute bill climbs either way.
SK Hynix and Sandisk Publish the First HBF Standard, Google and Tenstorrent Sign On
The company that won the HBM race just published the standard for the memory that comes after it.
At FMS 2026, SK Hynix and Sandisk released the first specification for High Bandwidth Flash, or HBF, a memory tier between HBM and the SSD: HBM-class bandwidth at the top grade, 0.4 to 3.0 TB/s, with the capacity and cost of NAND, up to 512GB per device. It connects over UCIe, so it sits in-package next to the accelerator, aimed at inference, where model weights and KV cache have outgrown what HBM can hold economically. Google and Tenstorrent have joined the consortium, and the spec is published openly through the Open Compute Project. SK Hynix also showed its 375-layer 4D NAND, with enterprise SSDs due early 2027.
Why this matters:
Inference is becoming a memory problem. A tier that keeps more model and context near the accelerator at flash prices hits the bottleneck that makes operators buy extra GPUs just for their HBM, the shortage SK hynix funded in the capital markets in Issue #114.
A hyperscaler and a merchant-silicon startup at the table is an indication of just how much impact the demand side is beginning to have on how the market operates, and who partners with whom.
SK Hynix and Sandisk want HBF to become the default the way HBM did, and locking in the open spec before Samsung and Micron back a rival is how you win a memory standard.
Google Locks In 155MW of Oklahoma Solar for Its Data Centres
One hyperscaler is solving the power problem the boring way, one contract at a time.
Google has signed a 15-year contract for the entire 155MW output of RWE's Crooked Creek solar project in McCurtain County, Oklahoma, due online in 2028 on the Southwest Power Pool grid. It is Google's second Oklahoma solar deal this year, after a 200MW agreement with Enlight in May, and it stacks on the $9 billion the company earmarked for cloud and AI infrastructure in the state last year. This is the unglamorous half of the build-out: lock up the power years before the racks arrive.
Why this matters:
A 15-year deal signed in 2026 for 2028 electrons is how you keep a data-centre roadmap on schedule when the interconnection queue runs years deep.
Renewables are still the fastest megawatts to build. Gas turbines and new reactors carry long lead times and longer politics; solar on an existing grid gets Google power by 2028, which is why hyperscalers keep signing it even while they chase nuclear headlines.
Oklahoma is quietly turning into a hyperscaler hub. Two PPAs and a $9bn commitment inside a year is how a new Virginia-style cluster starts, with the electrons lined up before the campuses fill.
SoftBank Lines Up an AI Data-Centre Push Across Southeast Asia
It is only a memorandum, but the names on it matter.
SB Telecom, a SoftBank subsidiary in Singapore, has signed a memorandum with SC Zeus, the data-centre arm of SC Capital Partners, and Malaysia's Robust HPC to build AI data centres across Malaysia, Thailand, Indonesia and Vietnam. The division of labour is the tell: SC Zeus sources land and builds and runs the sites, SB Telecom handles the fit-out, networking and connectivity, and Robust HPC procures and stands up the GPU clusters. No capacity, spend or timeline was disclosed, and an MoU commits no one to anything. What it marks is intent, the Stargate and Arm backer staking out Southeast Asia before the region's AI demand fully lands.
Why this matters:
Malaysia is already pulling in hyperscaler and neocloud money, and a SoftBank-badged consortium adds momentum to a region determined to build AI capacity at home.
Land and shells from a data-centre operator, connectivity from a telco, GPUs from an integrator: regional AI capacity now gets stood up without any one party owning the whole stack.
For now it is a memorandum without a megawatt attached. No figures, no sites, no dates, so treat it as a marker of direction, worth watching for whether it turns into steel and signed offtake.
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.