Key points
- Anthropic — Beyond the public offering, Nvidia, Microsoft, Google, TSMC, Broadcom, AMD and Micron have also committed a combined $26 billion of investment in Anthropic — which means the books of these chip and tech companies are now tied directly to its share price.
- Will — Will Wright, the designer of SimCity and The Sims, had been quiet for a long time since Spore's muted reception in 2008 and his subsequent departure from Maxis parent EA.
This is Salty Talk’s Daily AI Stew, brought to you by Robin.
Nvidia’s Q3 revenue doubles, board does not reinstate Jensen Huang
Start with Nvidia’s quarterly report — the numbers and the personnel question are two different things.
According to Yahoo Finance, Nvidia’s third-quarter revenue rose 103% year over year to $115.8 billion, beating market expectations, and the company also raised its guidance. This is the first quarterly report since founder and long-time chief executive Jensen Huang’s departure; the board says it will not reinstate him as CEO, and interim CEO Cress continues to run operations for now. The main growth in the earnings is still data center: revenue in that segment rose 110% year over year to $107.6 billion this quarter. Another thread is compliance: Chinese cloud providers that rent Nvidia’s advanced chips through intermediaries are being discussed with more risk attached, and after Huang’s departure, the related Asian partnerships — and the possibility of using intermediaries to sidestep US export restrictions on advanced chips — are under heavier scrutiny.
My read: the real story isn’t the growth rate but that decision-making authority at this company is temporarily unassigned. This quarter’s numbers prove demand didn’t turn around along with the personnel change, but who sets next quarter’s budget and who signs the supply commitments is, for now, an empty chair.
Anthropic opens at $111.50 on day one, market cap tops $2 trillion
Second item, from Wall Street: an AI company just completed its listing, and the market put a price on it on the spot.
Per reports from CNBC and Forbes, Anthropic’s opening price on its first trading day was $111.50; Forbes put the opening gain at 34%, and CNBC said its market cap crossed $2 trillion. According to Yahoo Finance, this is an artificial intelligence company founded in 2021; the IPO priced at $83 per share and rose to $111.50 once trading began. Beyond the public offering, Nvidia, Microsoft, Google, TSMC, Broadcom, AMD and Micron have also committed a combined $26 billion of investment in Anthropic — which means the books of these chip and tech companies are now tied directly to its share price.
My read: the opening price by itself says little. What has to be tracked is the first earnings report after listing: what the market bought today is its position in the compute supply chain, and whether that position converts into revenue still has to be settled by the numbers.
AI selloff pushes the S&P 500 and the Dow into correction territory
Now the market side: this leg down is no longer about a single company.
In the AI-driven selloff in tech stocks, the S&P 500 fell 3.7%, joining the Dow in correction territory — that is, at least 10% below a recent high; the Nasdaq gave back all of its 2026 gains and turned negative for the year. The selloff extends the previous session’s decline, when the S&P 500 fell 2.02% and the Dow fell 2.19%. Those hit most directly are investors holding AI and semiconductor stocks; for equity investors more broadly, what they see is the Nasdaq closing at least 10% below its record close on October 6, already inside a correction.
My read: where this decline lands is on index levels and financing costs, not on any one company’s first-day pop. Once the index as a whole enters correction territory, what conditions the data-center capex can borrow money on becomes a harder question than the share price.
US Senate passes national permitting reform bill, sends it to the House
On the policy line, one piece of legislation in the US Congress moved a step forward.
The US Senate passed legislation aimed at simplifying project permitting nationwide, covering energy and AI-related development. Note the boundaries of the reporting here: there is no vote count, no detail on the provisions, no implementation timetable, and no statement on whether it has become law. The bipartisan bill, named the Bipartisan American Affordability and Jobs Act, belongs to a long-running effort to streamline federal energy and infrastructure permitting; after Senate passage it goes to the House for consideration. If it is eventually enacted, the changes in the permitting process could speed up approvals for power transmission, energy projects and AI data centers, directly affecting developers’ schedules.
My read: this one can only be taken as procedural progress. The House still stands between it and enactment, so what really moves schedules is not the Senate vote but whether the provisions get amended later, and when enforcement begins.
Google lets Gemini operate a Mac on the user’s behalf
On the product line, Google pushed the assistant’s boundary out another notch.
Google is rolling out a Gemini capability that can perform actions on a Mac on the user’s behalf; Google’s blog frames it as taking repetitive chores off your hands, and the related headline also mentions background Mac task automation and a rollout starting in the US. Several gaps in the material are worth remembering: supported models, macOS versions and which tasks it can actually handle are all unstated, and access conditions, permission scope, privacy protections and safety limits are also unclear, so how much this feature actually covers can’t be judged yet. It marks a step for Gemini from conversational assistant toward an agentic one that controls the desktop, but the material doesn’t show whether the US-first rollout is general availability, a limited test, or subject to other conditions.
My read: the weight of this step is in permissions, not capability. Letting a model operate the desktop directly means granting it execution authority first and writing the rules afterward — which tasks are allowed, how mistaken actions get rolled back, how enterprise and personal accounts are tiered: none of that has an answer in the material.
Microsoft ships seven in-house AI models at Ignite 2026, betting on code cost-performance
On the same product line, Microsoft brought out seven models in one go this time, with the bet concentrated on writing code.
At Ignite 2026, Microsoft released seven in-house AI models at once, covering coding, image generation and editing, transcription and other tasks. The one drawing the most attention is the code-oriented MAI Code 2: Microsoft says its coding quality reaches 98% of GPT-6.1 Sol at roughly 40% of the price, and reaches 96% of Opus 5.5’s performance at just 5% of the price. Read these numbers as “the company says” — how the benchmarks are defined, how they were measured, what latency and stability look like: the company provided none of it, nor the full specs, open weights or licensing terms of the seven models. Microsoft also cited Artificial Analysis rankings and price comparisons to make its positioning, for instance calling MAI Image 2.7 the top image-editing performer and pitching Image 2.7 Flash at high-throughput generation. Reaction in the community splits two ways: some who have tried it think MAI Code 2’s coding experience really is comparable to Opus or Sol; others question whether the comparisons with GPT-6.1 Sol and Opus 5.5 are too bold, and ask why these models aren’t yet in Artificial Analysis’s index; still others argue that if Microsoft wants to grow share, it should offer open weights or open-source versions.
My read: with comparisons like these, ask first who defined the benchmark. Percentages only mean something on the same test and the same workload; the company gave no test method, so the cost-saving conclusion can only stand within its own framing.
Will Wright’s new simulation game will use AI to mine content from players’ memories
A different direction — a game-industry item: a veteran designer is back with a new title.
Will Wright, the designer of SimCity and The Sims, had been quiet for a long time since Spore’s muted reception in 2008 and his subsequent departure from Maxis parent EA. More than 15 years on, he is now promoting his next project, MyMems, planned for release next year: the game uses generative AI to build digital characters out of memories players provide, and those characters then team up to solve puzzles. His Gallium Studios and partner Artificial Agency said in a press release this morning that the new title differs from his earlier work by “turning simulation inward.” Wright offered two examples: “If my grandmother played a Grand Slam match against Serena Williams, who would win?” and “If I had to build a bridge over a short canyon, which three people from my life would I pick to help?”
My read: the first thing a product like this has to solve isn’t gameplay but consent. People who turn up in memories often never agreed to be made into characters, and how memories are stored, whether they can be deleted, and who carries the likeness and reputational risk for third parties will be asked about sooner than the game itself.
Google launches Gemini 4 Neon; community data shows stronger overall, weaker long-horizon reasoning
Back to the models themselves: Google has added another version to the lineup.
Google launched Gemini 4 Neon, positioned as a more capable new variant in the Gemini 4 family. The company gave no technical specifications, no pricing, no availability scope, and no verified benchmark results, so its real advantage over Carbon can’t be judged yet. Based on Artificial Analysis data, the community sees a split picture: Neon scores higher than Carbon overall, but lower on research-scale reasoning evaluations such as CritPt — meaning it leans toward broad capability, while Carbon remains aimed mainly at long-horizon, complex research tasks. Users also report that Neon runs at 127 tokens per second on OpenRouter (a token being the basic unit of text the model processes), against 46 for Carbon, which is an unverified community observation. One more point in the same index deserves a careful look: in the Xhigh test, Neon scores 61 on the intelligence index, above the 43 median for comparable models, but that evaluation round generated 283 million tokens, far above the 65.1 million median — there is compute spend behind the score. The naming itself is a burden too: with names like Neon and Carbon arriving so often, developers can hardly tell from a name what a model is actually good for; discussion of the OpenRouter speed gap also hasn’t reached consensus, it only shows that positioning and actual behavior aren’t aligned yet.
My read: read community benchmarks by use case. The one with the higher overall score isn’t necessarily right for long-horizon work, and there’s a generation volume sitting behind the score — if you’re really picking a model, run your own real workflow through it, then look at whether it’s expensive.
A claim that Claude took part in a Riemann hypothesis proof, with a denial the same day
A math-related item now: within the same daily digest, the same event got two accounts.
In its November 22, 2026 daily digest, AIHOT relayed that Anthropic mathematician Levent Alpöge announced on Friday that he and Claude had obtained a proof of the Riemann hypothesis and of the generalized Riemann hypothesis: the proof was released in the Lean language as 14 data packages, with a printed length expected to exceed 250,000 pages. In plain terms: the Riemann hypothesis says the real part of every non-trivial complex zero of the Riemann zeta function is 1/2, and the generalized version extends a similar conclusion to a family of Dirichlet L-functions; Lean is a theorem prover that can mechanically check proof steps, so writing an argument as code it can check helps expose logical gaps, but it can’t replace scrutiny of how the proposition was modeled or whether the underlying math library is correct. The report also said Stanford released a competing, purely human-written draft a few hours later. In the same day’s digest there was another report: Claude ultimately did not prove the Riemann hypothesis, and the actual contribution was raising the known lower bound on the proportion of zeros of the relevant zeta function from 41.6% to 67.2%. The 14 data packages and 250,000 pages describe the scale of the submitted material, not a conclusion the field has confirmed; read together, this can only count as an unverified claim for now — there is no original proof to read, no peer review, and no independent technical review.
My read: the lesson here is verifiability. What can be recomputed is a lower-bound number; what can’t be recomputed is the sentence “it has been proven” — until the formalization packages run and independent people review them, all an audience has is a claim.
Take these items together and both money and compute are still accelerating, but who gets to decide, and what the pricing standard is, hasn’t been settled: Nvidia’s board, Anthropic’s opening price, the index pushed into correction — they are really asking the same question. The market answered first; the rules are still catching up.
🎧 This episode is also available as a podcast: listen to Daily AI Stew · 2026-11-22.
