An AI lab is about to ask public markets for a hundred billion dollars.
At a two trillion dollar valuation, roughly double what private investors said four months ago. Much of the money is set to come from the companies that sell it the chips and the cloud it already runs on. The customer of everyone is about to sell equity to its suppliers.
Either the compute bill is a moat. Or the loop has started feeding on itself.
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.
Anthropic Picks Nasdaq for a $2 Trillion IPO and Wants Nvidia to Anchor It
Four months ago private investors marked Anthropic at $965 billion; it is now asking public markets for double that, with its chip supplier in the book.
Anthropic has selected Nasdaq for a listing reported at around $2 trillion, with the raise put at up to $100 billion. Nvidia is in talks to anchor it with up to $10 billion, though the reporting does not establish that as new money on top of its late-2025 commitment, and Amazon, already down for more than $100 billion of AWS spending over ten years, is expected to add equity. The business underneath is real: annualised revenue passed $65 billion in July, against $517 billion of compute contracted over eleven months (#123). Everything but the revenue and the May round is reported, not filed. The suppliers are being asked to fund their own demand.
Why this matters:
Doubling the mark in four months only works if the compute bill reads as a moat rather than a millstone. That is the whole bet.
Nvidia anchoring closes the loop it runs everywhere: it funds the lab, the lab buys its chips, and the purchases justify the valuation it just paid up for (#123).
A $65 billion run rate makes this a different listing from the cash-burners before it. If growth holds, $2 trillion is a floor; if it slips, it is the most exposed position in AI.
Z.ai Raises $5 Billion for Its Next GLM, and Names the Chips
Z.ai's $5 billion arrives with no valuation attached and one very deliberate line item: domestic-chip adaptation.
Z.ai, the Zhipu-built maker of the GLM models, completed around $5 billion of financing on 14 September: roughly $2 billion from a share placement, $3 billion in zero-coupon convertibles priced above the current mark. The money funds next-generation GLM foundation models, a self-training system, large-scale training and inference, and "domestic-chip adaptation and inference optimisation." It lands weeks after Zhipu's GLM-5.3 open weights (#123). No valuation was disclosed. China's model layer is not short of capital. It is short of NVIDIA, and spending to make that not matter.
Why this matters:
Premium-priced convertibles are how a listed growth stock raises, not a private lab. Capital is not the constraint here.
The earmark is the signal. Every dollar spent making GLM run on Chinese silicon is a dollar spent making NVIDIA optional, the wall #123 tracked through Huawei and CXMT.
A self-training system targets the human-label and data costs that pace frontier models. If it works, the open-to-closed gap closes on cost, not benchmarks.
Crusoe Signs Perplexity to Train and Serve on One Cloud, Size Undisclosed
Crusoe signed Perplexity's entire model lifecycle without disclosing a number, then two days later disclosed a very big one.
Crusoe and Perplexity signed a multi-year partnership on 15 September for Perplexity to run its full model lifecycle on Crusoe Cloud: training on dedicated NVIDIA GB300 NVL72 clusters over InfiniBand, serving through Crusoe's Managed Inference. Crusoe puts that inference at up to 9.9x faster time-to-first-token and 5x the throughput of vLLM, company-stated. No dollar figure, GPU count or megawatt was attached. It runs both ways: Crusoe is putting Perplexity Enterprise Pro and Max in front of its 1,800 staff. Two days later Crusoe raised $3.9 billion at a $30.9 billion valuation, roughly triple its October mark, with NVIDIA again among the backers.
Why this matters:
Training and inference on one vertically integrated stack is what separates Crusoe from GPU rental. For a product where latency is the experience, that is the whole pitch.
A full-lifecycle commitment nobody outside the deal can size is still an anchor. Perplexity routes across chips through its model-agnostic Computer, so read this as one leg of a spread book, not an exclusive.
Google's Crux took over Crusoe's paused 1.8GW Cheyenne campus only last week (#123). The company that builds AI factories for others is now filling its own.
TypeSafe Launches Jev, a Model That Returns Typed Decisions, Not Text
A model that cannot write a sentence claims to be two orders of magnitude cheaper than one that can, and nobody outside TypeSafe has checked.
TypeSafe launched Jev on 15 September, out of two years in stealth, the first of a class it calls System One models, named for Kahneman's fast thinking. Founder Diogo Almeida co-authored the InstructGPT work behind ChatGPT. Jev takes program state and typed questions and returns typed decisions: a choice of up to 255 options, a score, or a calibrated yes or no, each with a confidence score, computed in parallel. It claims 70 to 500 milliseconds a call, 20 to 200 times faster, input at $0.042 per million tokens with output free, and no hallucinations or type errors by construction. All company-stated, with no independent benchmark, no architecture paper and no public API. The pitch is not a smarter model. It is a cheaper unit of decision.
Why this matters:
Jev is a bet on the Jevons paradox its name invokes: make a decision cheap enough and demand explodes. It is the tokens-per-megawatt logic one rung up, where AI meets ordinary code.
If the valuable unit is a calibrated decision rather than a paragraph, the labs' edge in fluent text stops mattering across a large class of automation, and their margins go with it.
One reader has already charted Jev below Claude Sonnet 5 on TypeSafe's own example. Two orders of magnitude is a claim, not a result: it either reprices a tier of AI work or it is a fast demo.
Huawei Brings Its Ascend 960 Forward to Q1 2027 and Scales the Pod to 4,096 Cards
Huawei's next Ascend chip lands six months early, and the pod it was meant to fill arrived at a quarter the promised size.
At HUAWEI CONNECT 2026 on 17 September, Huawei said its Ascend 960DT will arrive in Q1 2027, roughly six months ahead of plan, and unveiled the Ascend 960 SuperPoD, linking up to 4,096 NPU cards for a company-stated 8 EFLOPS of FP8 and 1PB of HBM. It leans on near-packaged optics: Huawei's Hi-ONE modules, it says, cut more than 550kW from a large cluster's networking power against traditional 800G optics, at 99.8% availability. The Atlas 950 SuperCluster scales to 256,000 accelerator cards, and Huawei says more than 1,000 Ascend 910C supernodes are deployed. Analyst Rui Ma flagged the catch: the pod is far smaller than the 15,488-chip system Huawei first described. The performance is company-stated. The timing, a week before Trump and Xi meet, is not.
Why this matters:
This is the hardware end of the wall Z.ai is funding: a full stack that needs no US silicon. A 4,096-card pod is no Vera Rubin rack, but it ships.
Pulling a chip forward six months, a week before Trump and Xi meet, is a signal as much as a product. Export controls are buying time, not building a moat.
Early silicon with a shrunken system points at packaging and memory as the bottleneck, not the chip. Watch whether the pod grows to meet the chip, or the roadmap quietly meets the pod.
ByteDance's Profit Slips as It Builds Out Volcano Engine and Its Own Chips
ByteDance earned $20 billion in six months and still went backwards, because TikTok's ad money is now buying data centres.
ByteDance's first-half revenue reportedly reached $120 billion, up 30%, while net profit fell to $20 billion, the drop pinned on AI spending. The money is going into Volcano Engine, its AI cloud, which by one count holds close to half of China's model-as-a-service market, and into in-house inference chips meant to cut its reliance on NVIDIA. Bloomberg reported earlier this year that ByteDance was weighing up to $70 billion of 2026 capital spending, nearly triple 2025. Advertising and e-commerce still pay for it, but a consumer-internet company is turning into an infrastructure operator, buying and building the compute behind its Doubao assistant and Seedance video models. The profit line is the price: this is what a frontier cloud costs in China without a clear line to the best silicon.
Why this matters:
China's biggest AI buyer is funding a domestic cloud and its own silicon rather than betting on NVIDIA. Z.ai pays for the models, Huawei builds the chips, ByteDance supplies the demand.
Owning the model-as-a-service layer is hyperscaler leverage earned on tokens rather than ads, and it is the position AWS and Azure hold in the US.
Profit falling while revenue climbs 30% is a choice, and it holds only while TikTok pays the bill. The question is what the cloud and the chips return before the ad engine slows.
Bell Puts $52 Billion Behind a 1.2GW Saskatchewan Build, Tenant Chips Included
Bell is being credited with the largest private investment in Canadian history, though most of the $52.5 billion is someone else's money.
Bell is expanding its Saskatchewan AI campus to 1.2GW, from the 300MW already under construction, and calling the $52.5 billion total the single largest private capital investment in Canadian history. The figure does the lifting: it counts tenant computing equipment and partner-built power, not just Bell's land and buildings. Power is natural gas under Saskatchewan's bring-your-own-power framework, cooling is closed-loop with no municipal water, and the first phase of the original Sherwood project opens in 2027; the expansion depends on customer commitments and approvals. Announced on 14 September at the Canada Investment Summit, with Ottawa and the province branding it sovereign AI, it is a telco building a national compute base on gas. The headline is $52.5 billion. Bell's own cheque is a fraction of it.
Why this matters:
The tests that matter are Bell's own capex and how much of the 900MW is contracted. Neither was disclosed.
Gas under a bring-your-own-power framework is the pattern: like Google's Cheyenne and Duane Arnold deals (#123), the binding constraint is firm power on a timeline, and the grid cannot supply it fast enough.
Sovereign AI is doing political work here. It is how a telco's gas-fired campus becomes national infrastructure, clears permits and wins subsidy.
Everything Else
Planted raised $31.8 million, Google and Breakthrough Energy among backers, to scale robot-built behind-the-meter solar toward 1GW in 2027.
Branch Energy raised $33 million to bring behind-the-meter batteries to PJM, after Voltus lined up 100MW of storage for Google's data centres there.
SK Hynix is reportedly in talks with Intel to make DRAM and HBM in the US, on top of the $3.8 billion packaging plant it is building in Indiana.
Nvidia's B300 is renting near $7.40 a GPU-hour, a $1.40 premium over the B200 for 50% more memory and FP4 throughput, per Hyperstack.
BloombergNEF nearly doubled its forecast for US data-centre gas demand to 18 billion cubic feet a day by 2035, second only to LNG exports.
US grid congestion cost a record $17 billion in 2025, up 42% on the year, with PJM the worst at $3.2 billion.
A Pennsylvania regulator's study warned PJM's loss-of-load risk could run six to thirteen times its reliability target by 2030 if data-centre load keeps growing.
The US House passed the Ratepayer Protection Act 417-3, pushing states to make 100MW-plus loads cover their own grid costs instead of shifting them onto other customers.
UK regulators began design assessment of X-energy's 80MWe Xe-100 small reactor, the design behind Centrica's plan for up to 6GW of new British nuclear.
B2U switched on a 28MWh second-life battery in Texas and signed Waymo to supply retired EV packs, part of a pipeline above 1,000MWh.
Salesforce and Nvidia launched Koa, a CRM reasoning model post-trained on open-weight Nemotron and trained on synthetic data rather than customer records, claiming three times fewer errors on Salesforce's own benchmark.