Why weren't 163 million users enough to keep Character.AI alive?
Because it was paying about $10m a month for compute, and its best month ever brought in $5.3m. More users meant a bigger compute bill, not a profit. It filed for Chapter 11 on Monday.
Disney's opening bid is $20m, for a licence to some of the IP.
Popular and solvent turned out to be two different things.
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.
Character.AI Spent $56M on Compute in Eight Months, Then Filed for Chapter 11
Google paid $2.7 billion to license its models in 2024, and Disney is bidding $20 million for what is left.
Character Technologies, the company behind Character.AI, filed for Chapter 11 in Delaware on 12 October. Disney has made a $20 million stalking-horse bid for a non-exclusive, perpetual licence to some of its IP and technology, TechCrunch reports. Chief executive Colin Adams said the platform keeps running during the sale and shuts if no workable bid emerges. An affidavit Adams filed shows compute spending of $13.8 million a month in 2024 and $10 million a month in 2025, 404 Media reports. Compute cost $56.1 million in the first eight months of 2026, against $19.7 million on staff. Subscription revenue peaked at $2.8 million a month, and total revenue at $5.3 million. Google tops its creditor list, owed about $4.1 million, and DigitalOcean is second with about $2 million. The filings put current and past users at about 163 million.
Why this matters:
Usage was the cost. Every message burned compute, and subscriptions that peaked at $2.8 million a month never covered it.
Google took the founders and a licence to the models in 2024 for $2.7 billion. The company left behind kept the users and the compute bill.
OpenAI and Anthropic pay the same kind of bill, and it holds only while investors keep funding it. Anthropic's leaked prospectus shows $7.33 billion of compute against $4.6 billion of revenue in 2025, and the FT reports OpenAI expects $856 billion of compute through 2030.
CoreWeave's Credit Default Swaps Hit a Record as Its Bonds Trade Like Distressed Debt
CoreWeave borrowed at 2.875% a month ago. Its bonds now trade like it might not pay.
CoreWeave shares fell more than 15% on 14 October, SiliconANGLE reports, after investor Chamath Palihapitiya posted a chart of its $1.25 billion 2032 notes yielding almost 15%. DigitalOcean, TeraWulf and Hut 8 each fell more than 10%, and Oracle 8.8%. The next day CoreWeave's five-year credit default swaps reached a record 1,194.5 basis points intraday, Bloomberg reported, citing ICE Data Services. Its 9% notes due 2031 dropped to 79.9 cents on the dollar, a yield of 15.6%. In a memo confirmed by a spokeswoman, chief executive Mike Intrator told staff the market "acts irrationally" and that the backlog more than supports the company's debt. He blamed weak listings and debt deals at rivals for "bleeding into the narrative around CoreWeave". SiliconANGLE puts CoreWeave's outstanding obligations at about $50 billion, against a $104 billion backlog at the end of June. CoreWeave sold $4.2 billion of convertible notes at 2.875% on 22 September.
Why this matters:
Swaps above 1,000 basis points mark distressed credit. CoreWeave borrowed at 2.875% a month ago, and its swaps now imply a better than 60% chance of default within five years, Bloomberg says.
The backlog is the whole defence. Intrator says $104 billion of contracts covers about $50 billion of obligations, and bondholders doubt all of it pays in time.
The founders have already taken money off the table. Intrator, Brian Venturo and Brannin McBee have sold $3.3 billion of shares since the March 2025 listing, $2.4 billion of it this year, Bloomberg reports.
NVIDIA Is in Talks to Buy Reflection AI, a Week After Beam
NVIDIA already has $800 million in America's best-funded open-weight lab and may now take the rest, or its people.
NVIDIA is in early talks to buy Reflection AI or deepen its stake, Reuters reported on 10 October, citing the FT. A deal could come within weeks. One option is an acqui-hire, which would hire the staff and license the technology without taking control, and the FT says that would likely avoid a long regulatory review. NVIDIA led Reflection's $2 billion round in October 2025 and has put in about $800 million in total. In April, chief executive Misha Laskin said the latest round closed at a $25 billion pre-money valuation. Reflection released Beam, a 501 billion parameter open-weight model, five days before the report. SpaceX rented it the NVIDIA GB300 systems, in a deal worth up to $6.3 billion. Neither company has commented.
Why this matters:
NVIDIA keeps buying the demand for its own chips. It licensed Groq's technology in December and agreed to buy Hugging Face in September (#122).
The compute contracts are the loose end. Reflection owes SpaceX up to $6.3 billion and Nebius about $1 billion, and an acqui-hire leaves those contracts behind.
America's open-weight answer to DeepSeek would belong to its chip supplier. Beam trails Chinese models on most rows of its own benchmark table (#127).
Meta Ships Muse Wonder, Its First Frontier Model, and Keeps the Weights
Meta spent years giving its models away. Fourteen months after rebuilding its AI lab, its first frontier model comes through a paid API.
Meta released Muse Wonder and Muse Wonder Full on 12 October, the first frontier models from its Superintelligence Lab since Scale AI founder Alexandr Wang restructured its AI work, TechCrunch reports. Meta says both models score level with Fable 5.1 and GPT-6 Astra, and that Wonder beats Fable, Opus and Astra on MemoryBench, a test of long-term recall. Every score is Meta's own. Wonder costs $4 per million input tokens and $20 per million output tokens on the Meta Model API. Wonder Full is limited to Muse Code subscribers. Asked about open weights, Meta said they are coming for the smaller Muse Spark, but that the Wonder models are too big for many people to use and raise potential safety concerns. Meta trained both at Prometheus, its Ohio data centre that came online this year.
Why this matters:
Llama made Meta the best-known open-weight lab in the US. Its best model now sits behind an API, priced near Claude Opus 5.5, TechCrunch says.
Meta leads on MemoryBench because the other benchmarks are a tie. Recall across months means storing more context for every Muse user, and that cost lands on inference.
Prometheus now has a frontier model to show for it. Meta is selling access to it at the same time as it rents capacity from Nebius and CoreWeave.
Microsoft Bets Multiple Gigawatts on Its Own Maia Chips
Microsoft's next inference fleet runs on chips it designed, and it claims 40% more tokens per megawatt than GPUs.
Microsoft said on 12 October that it will scale its Maia chips to multi-gigawatt deployments from 2027 and make them the main inference system for its own products. Maia 300, the third generation, puts eight HBM4 stacks and 18TB/s of memory bandwidth on a 950W chip, with compute chiplets on TSMC's N2P process. Microsoft fits 96 of them in a 167kW rack. Microsoft claims about 40% more tokens per megawatt than leading GPU systems, and up to seven times Maia 200's throughput per rack. Neither figure is independent. Maia 200 is already in production, running Copilot and OpenAI models in the US, Europe and Australia. Reuters reports Microsoft could deploy more than a million Maia chips next year. Microsoft says the bottleneck is now "moving data efficiently, rather than raw compute power".
Why this matters:
Tokens per megawatt is the number that counts at this scale. Power caps every campus, and Microsoft is claiming 40% more output from the same grid connection.
The first workload to leave NVIDIA is Microsoft's own. Azure customers still rent GPUs, while Maia takes Copilot, GitHub Copilot and Foundry serving.
Every hyperscaler now has its own chip. Google has TPUs, Anthropic is leasing them through Broadcom (#126), Amazon has Trainium, and Microsoft now has Maia at gigawatt scale.
Amazon Is in Talks to Buy Decart for $7bn After Anthropic Walked Away
Amazon may pay $7 billion for software that makes NVIDIA-built models run on its own chips.
Amazon is negotiating a roughly $7 billion takeover of Decart, the Wall Street Journal reported on 11 October. Neither company has confirmed. Decart sells software that helps models run faster on NVIDIA, Google and Amazon chips, and licenses it to cloud providers and labs. It also builds real-time video models, Lucy and Oasis, and has moved them to Trainium. Amazon has said Decart ran real-time video four times faster on Trainium at half the cost of GPUs. Anthropic had discussed a $6 billion deal, Bloomberg reported in August, and walked away in September. NVIDIA joined Decart's $300 million round in May, which valued it near $4 billion.
Why this matters:
Trainium's problem is software, not silicon. Models written for NVIDIA need rework to move, and Decart's product removes much of it.
Amazon would own a tool its rivals use. Decart supports NVIDIA and Google chips today, and nobody has said whether that continues.
Microsoft is building its own chips and Amazon is shopping for the bridge to its own. Both moved against NVIDIA's inference share in the same week.
PJM Hits a Generation Alert Four Days After FERC Rejected Its Data Centre Plan
FERC threw out PJM's data centre curtailment plan on Friday, and by Monday PJM was short of supply.
PJM issued a maximum generation alert on 13 October, asking owners to defer maintenance as warm autumn weather met planned outages, T&D World reports. PJM then extended the alert into 14 October. The Department of Energy issued Order 202-26-50 the day before, letting PJM order specified plants to run and call on backup generators at large loads as a last resort before its highest emergency level. The order runs from 13 to 16 October across 13 states and DC. On 9 October, FERC rejected the core of PJM's large-load plan, Utility Dive reported. It threw out the scheme that would have let data centres connect early if they agreed to be curtailed first, along with the bring-your-own-capacity rules, saying both intrude on state jurisdiction. FERC approved a registry of every load of 50MW or more.
Why this matters:
The backstop is the data centres' own generators. DOE's order lets PJM run them before a top-level emergency, and their emissions do not count against rolling limits.
States now decide who gets cut. FERC Commissioner David Rosner wants non-critical data centres to go before homes, hospitals and schools.
Capacity prices stay higher for longer, Jefferies says. That suits Constellation, NRG, Talen and Vistra, and costs everyone else buying power in PJM.
Everything Else
DayOne signed a 15 billion ringgit loan, about $3.7 billion, with seven banks led by DBS and Standard Chartered for its Johor data centres.
Cipher Digital priced $850 million of zero-coupon convertible notes due 2032 at a 35% conversion premium, partly to fund its Barber Lake project.
Princeton Digital Group is seeking a $1.2 billion loan for a data centre in Johor, The Business Times reports.
SWI Group raised €800 million to turn Genesis Digital Assets' 15 crypto sites, about 1.3GW in the US and Sweden, into AI data centres.
Nebius signed a letter of intent for a large share of Volt's planned 600MW Rotterdam campus, whose first 200MW phase still needs about €7 billion, NL Times reports.
Salience Labs raised a $125 million Series B led by Bullhound, with AMD and Intel's venture arms joining, to build its photonic switch.
Infinigence AI raised a 2.3 billion yuan Series C and is preparing to list, TechNode reports.
Moonshot AI is close to a pre-IPO round valuing it at $50 billion, TMTPost reports, as it settles a shareholder dispute before a Hong Kong listing.
Croatia's Sisak-Moslavina County began rezoning for Pantheon, a planned 1GW, €50 billion AI campus targeting operations in early 2029.
NVIDIA committed $250 million of compute over five years to UK agentic safety research at Oxford, UCL and Imperial College London.
NVIDIA's share of China's AI accelerator shipments fell to 34% in the first half as Huawei reached 28%, Reuters reports, citing IDC.
NVIDIA gave its Vera Rubin rack and tray designs to the Open Compute Project, covering racks that draw 250kW.
Intel detailed its D-series inference GPU, formerly Crescent Island, with up to 480GB of LPDDR5X memory.