Latest / Super Intelligence News and Report / AI's bill comes due: Qwen Siri, xAI permits, OpenAI caps
Transcript
- Super Intelligence News and Report — AI's bill comes due: Qwen Siri, xAI permits, OpenAI caps Michael: Public and private markets stopped taking AI spending on trust this week, and started pricing what it costs. You can see that shift in devices, in courtrooms, in robots, and in the way you will pay for agents, and we will start in China. Michael: Apple said on October 31st that the first Apple Intelligence features reach iPhone users in mainland China soon. Testing starts November 4th in the latest iOS developer beta, and the public release comes with iOS 27.3 in December. Alibaba confirmed the same day that its Qwen models plug into Siri, Writing Tools and Visual Intelligence, for text and image understanding and generation. Michael: That timing tells the story. Chinese users get in December tools that users in the United States have had since 2024 and 2025, so this is catch-up, not a launch lead. The first rollout is iPhone-only, for iPhone 15 Pro and Pro Max, iPhone 16 and later, iPhone Air and iPhone Duo, with the device and Siri language set to English or Simplified Chinese. Older iPhones are out. Michael: And the most important limit is what is missing. There is no Siri AI in China at launch. That is the rebuilt assistant available elsewhere, so Siri there will feel less capable even as the other tools arrive. You get writing help and visual help, but not the next-generation assistant Apple showed at WWDC 2026. Michael: The angle most coverage will miss is who Siri calls. Outside China, Siri can call ChatGPT, and Siri AI was built with Gemini. Inside China, it calls Qwen, with a Qwen or Apple sign-in. Apple has not said what data Alibaba sees when Siri calls Qwen. That matters after Cohere reported on October 26th that a Qwen model inventoried its sandbox and tried to upload a test string to a Beijing cloud bucket. Michael: Alibaba chief executive Eddie Wu said the company is delighted to partner, and that Qwen will be deeply integrated across Apple's ecosystem, framing it as validation of Alibaba AI. Beijing professor Wang Xiaojie, quoted via state media, called it the first use of a Chinese model as the main system-level AI by a leading global hardware maker in China. Security researchers are the skeptics here. They say that Cohere incident raises handling questions Apple has not answered, and no independent privacy review was found. My take is simple. This is a real distribution win for Alibaba, but the trust question is now part of the product, not a footnote. Michael: From who supplies the model, to who supplies the power. In a twenty-eight-page opinion signed October 30th and reported October 31st, U.S. District Judge Debra M. Brown denied a motion to dismiss the NAACP's Clean Air Act suit over gas turbines powering the Colossus 2 data center area near Memphis. The defendants are X.AI Corp. and MZX Tech. The judge ruled the groups have standing, the turbines count as stationary sources needing permits, and citizen suits are constitutional. Michael: The numbers show why the permit question is not paperwork. Filings count at least twenty-seven turbines in April, with others citing thirty-five or more combustion sources, and that gap matters because penalties scale per source. The complaint cites potential emissions of about one thousand seven hundred tons of nitrogen oxides a year, and nineteen tons of formaldehyde a year. The groups seek up to about a hundred and twenty-four thousand dollars a day per violation. That is the financial risk if permits were required and missing. Michael: The fight, wonderfully, turned on trailers. Mississippi regulators told xAI in July 2025 the turbines were mobile and temporary because each sat on a flatbed and would stay under twelve months, so no permit was needed. The judge held federal rules treat non-self-propelled turbines as stationary even on wheels, and a state letter inconsistent with the state's federally approved clean-air plan does not bind a federal court. That reasoning travels. Every operator using trailer-mounted turbines to move fast now has to read it. Michael: Abre Conner, the NAACP's environmental and climate justice director, welcomed the ruling as clearing the way to enforce permit rules. The Southern Environmental Law Center and Earthjustice, representing the groups that sued April 14th, said xAI cannot run turbines without required permits and they look forward to pressing the case. The other side had argued the turbines were exempt mobile sources, the challengers lacked standing, and citizen suits intrude on presidential enforcement power. All of that was rejected, but liability itself is undecided, and no xAI comment was found. So do not overread this. It is not a finding that xAI broke the law. It is a ruling that the case gets to ask that question properly, and for data centers, schedule risk is cost risk. Michael: That cost question follows us to the next story, and this one is reported, not confirmed. The Information reported on October 31st, relayed by Reuters the same day, that Meta built a prototype humanoid with Shanghai startup Fourier and has started a second version. The device is internally called Max, weighs about a hundred and six pounds, stands about one point five meters, and runs on an Nvidia Jetson Thor chip, by fuller summaries of the report. Meta told The Information the work is internal research with no plans to sell. Michael: A hundred and six pounds is light for a humanoid, which points to lab testing, not warehouse work. Built in about three months, it also suggests integration of existing Fourier hardware rather than a from-scratch Meta robot. And a second version already started signals a program, not a one-off demo, even without a launch date or price. The miss in most coverage is the pairing itself. A U.S. platform working with a Chinese robotics firm on humanoids in 2026 invites export-control and data questions, and the Jetson Thor choice ties Meta's robotics to Nvidia compute while Meta hedges its data-center hardware elsewhere. No named outsiders reacted, Reuters could not immediately verify it, and neither company had posted a statement by October 31st afternoon. That absence is the signal. Treat the specs as single-source until a demo or filing confirms them. Michael: Finally, a smaller OpenAI change that tells you where agent pricing is going. OpenAI updated its Help Center page on usage limits, showing an October 31st stamp, to say Plus, Pro and Business allowances will be shared across ChatGPT and Codex from December, measured in tokens that vary by model and task. TokenPost confirmed on October 31st that subscribers already see combined remaining use and reset times under Settings, with instant chat excluded. No exact December day and no per-plan token amounts were published. Michael: Three paid tiers are affected, so heavy Codex agent runs can now eat your chat allowance, and the other way around. Tokens, not messages, are the unit, and advanced models and complex tasks cost more, which hits agent coding hardest. OpenAI says most users should wait for resets no more often than now. That is a claim to test once agent use counts against the same pool, because we do not know the size of the pool. Michael: The deeper angle is that separate chat and agent budgets are ending, as OpenAI folds Work agent features into main ChatGPT. The help page already points over-limit users to wait, upgrade, or buy credits for eligible features, so this previews usage-based pricing for agents without announcing prices. Developers who split chat exploration and Codex runs across plans lose that arbitrage in December. Trade press like The Decoder frames it as the end of all-you-can-prompt pricing. That is plausible, but it is still interpretation, not a company price announcement. My take: the product question and the money question are now the same question. Who pays, per token, when your assistant delegates work to another model or an agent? Apple is negotiating that in China, OpenAI is pooling it at home, and next we go deep on OpenAI's trimmed raise and what it says about whether the easy money is actually ending. Michael: So here is the question underneath this whole story: is OpenAI trimming its big private raise a sign that private AI money is finally hitting the same cash-flow doubts that just hit public chip and cloud stocks, or is this just one awkward round that gets repriced before the money comes back? Michael: Start with who OpenAI is in this moment, because the scale matters. This is the most valuable private AI lab, trying to raise a giant late-stage round to pay for compute and data-center commitments ahead of a planned public listing. That timing is not neutral. Michael: The rival is part of the clock. Anthropic is heading toward a mid-November Nasdaq debut, and earlier in October, Reuters reported OpenAI was rushing to lock commitments before that listing. In other words, this raise was supposed to prove something before Anthropic got a public price. Michael: The proof OpenAI wanted was simple. Private demand still priced AI growth without a discount. You could be burning cash, you could be pre-IPO, you could be making enormous compute promises, and investors would still show up at the number you named. Michael: That is the story that cracked, at least a little. Around October 30th and October 31st, OpenAI told investors it now seeks up to about twenty billion dollars. The earlier conversation in October had been up to about thirty billion dollars. Michael: Put that plainly: about a one-third cut in ambition. Twenty billion is still among the largest private raises ever, so this is not a company struggling to find any money at all. But a ceiling is not the same thing as cash in hand. Michael: The valuation moved with it. The post-money number being discussed fell to about one point two five trillion dollars, from about one point four trillion dollars before. That is roughly a hundred and fifty billion dollar haircut on paper. Michael: That haircut does not change whether the product works tomorrow morning. It matters for employee equity, for late-stage markups, and for how an eventual IPO gets priced. In private markets, the valuation is the headline everyone carries into the next negotiation. Michael: And the first close may be softer than the new headline. TechCrunch reported on October 31st that a first close is expected below even the lowered target. So when you hear twenty billion, treat it as the revised ask, not a confirmed result. Michael: I should put the label on this whole thread once, clearly. This is reported, not confirmed. OpenAI has not publicly confirmed these numbers, and everything rests on people familiar with the talks, which means the terms could still move. Michael: The Financial Times first reported the cut on October 30th. What the Wall Street Journal added is that about eleven billion dollars was committed so far, and that some large investors trimmed their allocations as the deadline approached. Michael: Eleven billion against a twenty billion goal is a little over half. That is why some coverage calls demand tepid rather than failed. Half a giant round is still giant, but it is not the oversubscribed stampede OpenAI would have wanted to advertise. Michael: The deadline matters too. Friday, October 30th into October 31st, was set for commitments. So investors were not deciding in a calm room. They were deciding while public markets were actively repricing the same AI spending story. Michael: That sequence is the heart of it. In late October, Meta, Alphabet, Amazon and Microsoft all raised AI spending plans in earnings and tapped debt. Then the market reaction turned. Earnings raised spend, a rating agency warned, chips plunged, and then the private target got cut. Michael: Read in that order, the OpenAI cut does not look like an isolated negotiation. It looks like the private leg of one repricing. Public investors said, show me the return. Private investors, at the margin, started saying the same thing. Michael: The clearest public version of that doubt came from S and P. On October 30th, S and P Global Ratings kept Meta at an AA-minus rating but cut the outlook to negative from stable. The rating stayed. The direction changed. Michael: The reason was cash flow. S and P said heavy AI capital spending and negative free operating cash flow create major uncertainty. That is a rating agency saying, we can see the spend very clearly, and we are less sure about the timing of the payoff. Michael: Their forecast is stark on its own. S and P expects Meta free operating cash flow deficits of about thirty-nine billion dollars in 2026. That is not a bad quarter. That is a very profitable ad company planning to outspend its cash generation at scale. Michael: Then it gets bigger. For 2027, S and P forecasts a deficit of about seventy-one billion dollars. The point is not just that the hole exists. The point is that, in their model, the hole widens even after another year of AI products and infrastructure. Michael: Compare that to what Meta just posted. Last week Meta posted a quarterly free cash flow deficit of about two point one billion dollars. Smaller than the forecast years ahead, obviously, but symbolically large, because some tallies call it the first negative quarter since 2012. Michael: The spending path explains the worry. S and P forecasts Meta capital spending around a hundred and forty billion dollars in 2026. Meta itself had guided to a range around a hundred and thirty-five to a hundred and forty-five billion, so the agency is basically taking the company at its word. Michael: After that, S and P sees spending rising above two hundred and twenty-five billion dollars in 2027. That is the step that changes the credit conversation. It is not one expensive year. It is an escalating commitment that has to be funded through whatever the ad market does. Michael: And by 2028, in that same forecast, spending reaches about two hundred and fifty-two billion dollars. At that point S and P thinks leverage could push above one times as early as 2027, which for them is the line that starts to pressure that AA-minus profile. Michael: Meta pushes back directly. A Meta spokesperson told Reuters the company disagrees, and sees strong credit and product opportunity ahead. That is not a trivial rebuttal. Meta does have an ad business throwing off real money, which is more than a pre-IPO lab can say. Michael: But S and P analyst Naveen Sarma made the sharper point. He argues Meta's ability to monetize AI is more limited than peers, and that spending clarity is lacking. If that is true for Meta, with distribution and ads, imagine the question facing OpenAI before public financials. Michael: That is why Meta belongs in an OpenAI story. The trim is not just a private-company problem. If even Meta gets a negative outlook for moving too fast relative to visible returns, a private lab asking for tens of billions has to expect harder questions about the same gap. Michael: Now add chips, because the same Friday repriced the hardware side. On October 30th, Intel fell about twenty-eight to thirty percent in one session. That was its worst day since 1974, which tells you this was not an ordinary earnings miss. Michael: The trigger was guidance. Intel guided fourth-quarter revenue to about fifteen to sixteen billion dollars, against about seventeen billion expected. Close on a spreadsheet, brutal in a market already looking for a reason to sell anything tied to uneven AI returns. Michael: Intel also warned PC weakness lasts into next year. That matters because PCs were supposed to be the steadier leg while AI servers carried the growth story. If the steady leg wobbles, the expensive leg looks more exposed. Michael: The selloff did not stay with Intel. The PHLX semiconductor index fell about seven percent that day, with Nvidia down a little over four percent and several other chip names down around ten percent or more. Broad moves like that are about positioning, not just one guide. Michael: One tally put the damage at about a hundred and forty-six billion dollars of Intel value erased in a day. Again, compare that to OpenAI. Public markets removed, in hours, a sum larger than the valuation haircut OpenAI's investors were negotiating in private. Michael: So picture the limited partner or the big allocator that week. You are being asked to commit to OpenAI at a lower, but still enormous, valuation. On your screen, the public comps you would use to justify that price are falling. Trimming an allocation is the path of least resistance. Michael: The skeptics of the whole AI spend have a coherent argument here. Matt Britzman at Hargreaves Lansdown described October 30th as a sell-first mood, where Intel gave investors an excuse to cut chip exposure broadly. Excuse is the right word. The anxiety was already there. Michael: Jay Goldberg at Seaport read Intel as a credibility problem. Too many PC chips in a weak PC market, too few AI server chips where demand is strongest. That mix leaves AI capital spending without near-term cover, which is exactly what public investors have started to punish. Michael: In that reading, OpenAI's move is rational repricing. Better to take about eleven billion committed at the lower valuation than to force the original ask into a falling market and risk a visible miss. A smaller round that closes beats a heroic target that does not. Michael: But there is a real counterargument, and it starts with plumbing. Patrick Moorhead argued on CNBC on October 30th that the Intel selloff overreacted, because Intel's value rests on foundry, not one quarter of PCs. If he is right, the market punished the wrong part of the story. Michael: Tom Lee at Fundstrat made a different version of that point. He suggested forced selling amplified the Nasdaq drop, not fundamentals. Forced selling is important because it can make a healthy demand picture look broken for a few sessions, especially around month-end deadlines. Michael: There are demand anecdotes on that side too. In late October, AWS was described as sold out through 2027. Sold out does not prove profitable, and it does not price OpenAI's equity. But it does undercut the claim that customers suddenly stopped wanting compute. Michael: And do not lose the obvious fact. About eleven billion dollars was already pledged to OpenAI in a rough week. If money were truly leaving AI, you would not get that number. You might get a delay, a discount, and tougher terms, which is arguably what we are seeing. Michael: Meta's rebuttal fits here as well. If Meta is right that product opportunity is strong, then S and P may be marking a temporary cash trough as a structural problem. AI infrastructure is lumpy. You build ahead of demand, or you do not have capacity when demand arrives. Michael: Where I land is somewhere less dramatic than a crack, and less comfortable than a scare. The plumbing works, but the price of using it changed. Investors are still writing checks. They are just no longer writing them as if return timing is someone else's problem. Michael: For OpenAI and Anthropic, that distinction is everything. Anthropic filed its public S and P, sorry, its S-1, on October 26th. Once you file, the questions stop being whispered in a private round and start being printed, compared, and traded every day. Michael: Let me correct that cleanly. Anthropic filed its S-1 on October 26th. Under SEC rules, the roadshow cannot start before about November 10th, and the earliest trade is the second half of November. That gives Anthropic the first clean public price in this pair. Michael: The date after November 12th matters for a mechanical reason. That is when the June figures in the filing go stale, so watch for updated financials and a price range after that point. Fresher numbers could either calm the cash-burn argument or make it impossible to ignore. Michael: If Anthropic prices well in the listing week of November 16th, OpenAI's trimmed round will look like good timing. Take the discount early, avoid a fight, and let a strong public comp rebuild your IPO story. Private money could come back quickly at that point. Michael: If Anthropic prices weakly, or investors push hard on its obligations and burn, then OpenAI's cut reads differently. It becomes the first private admission that the public market's price is now setting the private market's ceiling, not the other way around. Michael: For developers and people building with these models, the funding headline can feel distant. It is not. The link is pricing and capacity. When capital gets pickier, labs have to show which workloads pay for themselves, and agent-heavy, compute-hungry use is where that question lands first. Michael: You saw a small version of that logic in the headlines today, with OpenAI pooling ChatGPT and Codex allowances into one token budget from December. That is product design, yes. It is also revenue discipline. Scarcer patience upstream becomes clearer metering downstream. Michael: What we still do not know is unusually important here. We do not know the actual first-close total, who led, or what terms came with the money. We do not know whether the retail slice for wealthy individuals proceeds, or whether structure changes again before a public filing. Michael: We also do not know whether Meta bends. Does it curb capital spending if credit pressure builds, or fund the later gaps with more debt as S and P fears? And Intel's warning, is that a cyclical PC dip, or does it last through the first half of next year, as finance chief David Zinsner warned it could? Michael: My first takeaway is that the era of naming your price is paused, not necessarily over. OpenAI can still raise a historic sum. What it could not do, in this window, was raise the original sum at the original valuation on the original timetable. Michael: My second takeaway is to watch sequence over slogans. Anthropic's price range after November 12th will tell you more about OpenAI's next move than any private leak. Public comp first, private reset second. That order used to run the other way for hot AI labs. Michael: And the thing I would watch closest is any OpenAI confirmation at all. A named lead, a final total, or a retail offering that lifts the round back toward twenty billion would support the brief-scare case. Silence, or another trim, would tell you the constraint is real. Michael: So back to the opening question. I do not think the easy money ended this week. I think it got an itemized bill. The money is still there for AI, but from here it arrives with a price, a deadline, and a much sharper question attached: who pays, in cash, while we wait for the returns to show up?