TODAY IN AI
● Anthropic holds its pre-IPO investor day today at its San Francisco headquarters, with a reported target of $1.8–2 trillion and trading aimed to begin before Thanksgiving.
● Microsoft detailed Maia 300: 288GB of HBM4, 18 TB/s of memory bandwidth and 13.8 PFLOPS FP4, built on TSMC 2nm. Deployment starts 2027.
● Linus Torvalds shipped Linux kernel patches written with Claude Code and measured roughly a 3x gain on one cached-read workload. They are experimental and not merged.
● IonQ is buying Elisity in a stock deal worth up to $1.3 billion, and TSMC is expected to post a sixth consecutive record quarter.
Anthropic opens the books to investors today
Anthropic is holding a pre-IPO investor day at its San Francisco headquarters today, 14 October, where invited institutional investors can question executives directly. Bloomberg reported the plan, citing informed sources.
The valuation figure circulating among potential investors is $1.8 to $2 trillion. Official marketing could begin as early as the week of 9 November, with trading targeted to start before Thanksgiving on 26 November. Those plans can still move.
Two bits of context worth holding alongside the number. OpenAI has reportedly delayed its own listing, having previously been reported as planning to raise at least $30 billion at roughly a $1.4 trillion valuation. And US IPOs this year, excluding SpaceX and SK Hynix, have a weighted average return of about -4%, against roughly +12% for the S&P 500 and +20% for the Nasdaq 100. Anthropic would be listing into a market that has not rewarded new issues.
Full detail: Anthropic investor day, valuation target and timeline.
Microsoft Maia 300 — the specs
Microsoft unveiled its third-generation in-house accelerator on 12 October. The headline numbers:
| Spec | Maia 300 |
|---|
| Compute | 13.8 PFLOPS FP4 / 6.9 PFLOPS FP8 |
| Memory | 288 GB HBM4, eight 36 GB stacks |
| Memory bandwidth | 18 TB/s |
| Process | TSMC N2P (2 nm) main compute die |
| Power | 950 W SoC TDP (Maia 200: 750 W) |
| Scale-up interconnect | 3.2 TB/s bidirectional |
| Rack density | 96 accelerators; clusters near 12,000 |
| Deployment | From 2027 |
Microsoft claims up to 40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale throughput of Maia 200 in a fully populated rack. Treat the first claim carefully — Microsoft has not specified the competing GPU configurations it measured against, so it cannot be independently checked. Reuters reports Microsoft could deploy more than a million of these chips next year.
Full specs: Microsoft Maia 300 breakdown.
Torvalds wrote kernel patches with Claude Code
On 10 October Linus Torvalds committed eight experimental patches titled "mm: read from the page cache without folio references" to a branch named noref-filemap-read-with-llm, based on Linux 7.3-rc6. The code, the tests, the review and the first draft of the commit message were produced with Claude Code under his direction. He ran the benchmarks himself.
On eight threads reading 16 bytes from the same cached file, throughput went from about 10 million reads per second to 30–33 million. Four threads on separate cores went from 9–12 million to 25–26 million. A single thread barely moved — 8 million to nearly 9 million.
That threefold figure applies to one contended workload, not to Linux generally, and Torvalds lists unresolved problems himself: memory allocation behaviour, compatibility, possible regressions, and the cost of delaying memory reuse. The patches are not in mainline.
Detail and the full numbers: Torvalds, Claude Code and the Linux kernel.
Also today
IonQ is acquiring Elisity. The quantum computing company agreed on 12 October to buy the Zero Trust security firm in a stock deal worth up to $1.3 billion, including up to $150 million contingent on performance. Closing is expected in Q4 2026, subject to approvals.
TSMC heads for a sixth straight record quarter. Analyst estimates collected by Reuters put third-quarter net profit near NT$746.3 billion, roughly $23.4 billion, up about 65% year on year, driven by AI chip demand and CoWoS packaging capacity. These are estimates ahead of the official report, not results.
Ambi raised $26 million. Founded by a former Plaud executive, the company ships apps that can optionally listen continuously on phone, Mac and Apple Watch, so its agent can suggest things during meetings and start follow-up work afterwards. Whatever you make of the premise, always-on capture is becoming a funded category rather than a thought experiment.
DigitalOcean launched two agent services. Agent Data Services, a managed data layer for agent context, and Agent Learning Services, covering Insights, Signals, Evaluations and Simulations for capturing agent traces, spotting failure modes and testing fixes. Both entered public or private preview on 12 October. No pricing published.
Brad Smith called for an international AI safety framework. The Microsoft president argued on 12 October that AI carries lighter oversight than ordinary consumer products, proposing independent evaluations, mandatory incident reporting, notification for people affected by AI agents, and an emergency shutdown mechanism. Proposals, not rules.
The UK AI Safety Institute says it now has access to frontier models. AI Minister Kanishka Narayan said on 12 October that cooperation with the US government and developers including Google has restored access, and that the institute is evaluating systems including Gemini 4 Argon. Narayan also said the government will consider stronger legal protections against advanced AI risks, including preparing for serious incidents involving autonomous agents and stricter near-miss reporting.
European AI Innovation Month starts today, running to 17 November and showcasing European AI projects across business, research and the public sector.
Sources
Notebookcheck — Maia 300 specifications ·
Linuxiac — Torvalds kernel patches ·
UPI — Microsoft announces Maia 300 ·
AIdapted — 13 October roundup