Opening

I'm Muyan, and this is Mark's Tech Insights. Today's biggest bombshell: Anthropic wrote the line "the company may be approaching the recursive self-improvement threshold" into the risk factors section of its IPO prospectus. The same document contains another eye-watering set of numbers: a single quarter's revenue against purchase commitments approaching the trillion-dollar range. I'll explain how to read that gap in a bit. Also on the docket: Apple pulling 19 deepfake tools disguised as pet apps, OpenAI's funding round shrinking while its valuation holds, and a whole wave of sovereign AI money.

Pre-recording editorial note (do not read aloud): For items 3 through 7 below, all we currently have are aggregator-style weekly roundups and category pages. Editors should fill in the individual permalinks before recording. For figures marked "pending official confirmation," the host should keep the hedge verbally.

---

Today's Top Stories

1. Anthropic warns: the company is already close to the "AI building AI" RSI threshold

• Source: TechCrunch (https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/)
• Summary: In an S-1/A amendment filed November 2, Anthropic added a risk disclosure stating that internal estimates suggest the company may be "approaching" its own self-defined recursive self-improvement threshold. In plain language, the threshold roughly amounts to this: if a model can compress into a single year the algorithmic progress that previously took several years, the line has been crossed (for the exact year ranges, refer to the prospectus text; I'm not quoting it verbatim here). The company also stressed that no model has crossed the line to date, and that it has no plans to let Claude design its own successor unsupervised.
• Most surprising bit: According to TechCrunch's reporting, Anthropic admitted it had previously used AI to build a new model without realizing at the time that capabilities had jumped, meaning the "level-up" was only discovered after the fact.
• Taiwan angle: Taiwanese teams are now using AI for coding and model tuning almost across the board, but internal evaluations mostly measure "did the task get completed," rarely "how much stronger is this version than the last, and along which dimensions." Without a mechanism to measure capability jumps, you'll end up exactly like Anthropic, finding out after the fact.
• Discussion points:
  1. Writing extreme risks into the "risk factors" section of an S-1 is actually standard liability-shielding practice in US IPOs. Is this an internal alarm bell, or lawyers doing what lawyers do?
  2. If capability jumps are only detected retroactively, how should the "pre-release evaluation" process be redesigned?
  3. The problem with self-defined thresholds is that the party setting them is the same party crossing them. Who verifies?
• Script suggestion: "Let me offer a balanced take first: writing the most dramatic risks into a prospectus is entirely common in US listing practice. The lawyer's logic is, 'I disclosed it, so you can't later say I didn't.' So don't immediately read this as an internal red alert. But I'm still leading with it, for one reason: they admitted they once used AI to build a new model and didn't notice at the time that its capabilities had jumped a level. That's not a prediction, it's a retrospective. And what makes retrospectives scarier than predictions is that they show our existing observation tools can't keep pace with how fast the system itself is changing. That's the same warning for every team in Taiwan doing internal model evaluation."

---

2. WIRED investigation: Apple pulls 19 disguised apps, exposing a deepfake production line

• Source: Original WIRED investigation, 2026/11/02 (individual permalink pending; search keywords: WIRED AI apps deepfake investigation); Chinese-language syndication reference: TUAW (https://www.tuaw.com/2026/11/02/apple-removes-19-ai-apps-hiding-abusive-deepfake-tools)
• Summary: WIRED traced over 400 interlinked apps and domains presenting themselves as pet-photo generators, language courses, and AI fitness coaches, while hiding adult and violent deepfake generation features behind server-side switches. The method: pass review cleanly, then "transform" after going live. Apple has removed 19 apps (including the pet portrait app Viefur), Meta says it acted against tens of thousands of ad accounts and pages, and Google removed some apps without disclosing an exact number.
• Most surprising bit: This isn't a handful of bad developers, it's an organized commercial production line that grows its own traffic and cross-promotes internally. Researchers' conclusion: enforcement has never caught up.
• Taiwan angle: Taiwan still lacks any equivalent "continuous platform oversight" tooling, meaning apps that quietly swap in new functionality post-launch face essentially no routine tracking. The draft AI Basic Act is also still under discussion. The more immediate impact on local developers is that review will only get stricter, and innocent small utility apps will take collateral damage.
• Discussion points:
  1. Server-side switches are legitimate software engineering practice. How should review systems distinguish feature flags from "post-launch transformation"?
  2. Taking down 19 apps out of a network with 400+ nodes, is that a flesh wound or a real blow?
  3. Remedies for victims in Taiwan are essentially nonexistent, and that gap is harder to fix than platform liability.
• Script suggestion: "What stings most about this one isn't that 'someone's doing something bad', it's the commercial maturity of it. Some people build the clean shell apps to pass review, some buy ad accounts to drive traffic, some run the backend switches. The division of labor is as clean as a normal company's. By contrast, platform review is still stuck on static checks at the moment of submission, and that time gap is exactly where the whole production line makes its money. Taiwan's position is even more awkward: we don't even have post-launch continuous monitoring tools yet, and the draft AI Basic Act is still being debated. So next time you find a free, genuinely useful AI photo editing app, seriously, go check what else that developer has shipped."

---

3. Cohere closes $5.01B Series E, with Prosperity7 and Nvidia leading a sovereign enterprise AI bet

• Source: Parsers Funding Rounds Report weekly, 2026/11/03 (discovery source, aggregator-style; individual article or Cohere official press release link pending) https://parsers.substack.com/p/funding-rounds-report-weekly-of-november
• Summary: In the same week everyone was shouting about an AI bubble bursting, Canadian enterprise- and government-focused AI company Cohere reportedly landed a $5.01 billion Series E, with Nvidia appearing on the investor list again. Reports also mention progress this year on a merger with Germany's Aleph Alpha, though that item currently rests on a single aggregator source and awaits official confirmation.
• Most surprising bit: The biggest check isn't going to consumer chatbots, it's going toward "helping countries and enterprises run models behind closed doors."
• Taiwan angle: Taiwan's sovereign AI conversation in recent years has mostly centered on compute procurement and models like TAIDE. But Cohere proves something else: the commercial core of sovereign AI is deployment and compliance, not benchmark scores. What Taiwan actually lacks is the systems integration capability to fit models into the regulatory workflows of finance, healthcare, and the public sector.
• Discussion points:
  1. The "non-OpenAI camp" is consolidating. Is that risk diversification, or just a different flavor of concentration?
  2. Nvidia invests in customers while selling cards to those same customers. How should we account for this circular investment?
  3. If the Cohere–Aleph Alpha merger news holds up, what does it mean for Europe's sovereign AI landscape?
• Script suggestion: "Caveat first: on the Cohere–Aleph Alpha merger, I've only got one aggregator outlet reporting it, so I'm holding back until there's an official announcement. But the $5.01 billion figure alone is noteworthy, because it runs completely against market sentiment. Everyone's shouting 'bubble' while money flows toward 'inside the enterprise, data stays onshore.' My read is that the consumer chatbot story has run its course, and the next round of competition is about who can pass a bank's security audit and who can deploy inside a government data center. That's actually good news for Taiwan, because that's a systems integration strength, not a model-parameter race."

---

4. UK's Fractile raises ~$700Mata700M\; at\; a$6.5B valuation, betting on SRAM inference chips

• Source: AI Weekly funding tracker (discovery source, aggregator-style; Fractile official announcement or individual Sifted/FT link pending) https://aiweekly.co/
• Summary: UK startup Fractile is taking on Nvidia head-on with SRAM-based inference chips, reportedly raising roughly $700millionata700\; million\; at\; a$6.5 billion valuation, backed by the UK's sovereign AI fund. Reports additionally mention the money supports an inference chip order from Anthropic worth roughly $250 million, again, a single-source claim, so I'm downgrading it to "reportedly."
• Most surprising bit: If the order claim holds, it means model companies are now funding Nvidia's challengers out of their own pockets, customers becoming their rival's financiers.
• Taiwan angle: SRAM-centric inference architectures rely on large on-die SRAM rather than attached HBM, which directly reshapes packaging requirements. What's worth watching for Taiwanese suppliers: these designs don't necessarily have to go through the most congested CoWoS path, and may instead shift volume toward other advanced packaging routes and test. In other words, this isn't necessarily just "one more order", it could be a reshuffling of capacity allocation.
• Discussion points:
  1. For inference chips without HBM, how do you trade off memory capacity ceilings against model size?
  2. Sovereign funds investing in chip startups: industrial policy, or a national-level bet?
  3. From design tapeout to volume production, how long is the timeline Fractile actually faces?
• Script suggestion: "I want to spell out the technical point here. Nvidia's path is compute die plus attached HBM, which is why CoWoS has been the bottleneck. Fractile's approach is to cram as much memory as possible onto the die, trading SRAM for bandwidth. The biggest constraint on that path is capacity, but the upside is that the whole supply chain doesn't have to fight the entire world for the same packaging capacity. So for Taiwan this isn't just 'another customer', if this architecture really scales, advanced packaging and test scheduling get reshuffled. Of course, there's a long road from design to tapeout to actual shipment, so I'd treat this as a variable worth tracking, not a conclusion."

---

5. South Korea's FY2027 budget elevates AI to national security status, including KRW 4.7T for homegrown frontier models

• Source: AI Weekly daily brief (discovery source, aggregator-style; South Korean Ministry of Economy and Finance budget press release or individual Yonhap link pending) https://aiweekly.co/ai-news-today
• Summary: South Korea's FY2027 budget treats AI as a national-security-grade investment, reportedly totaling around $16 billion. However, only three line items are currently visible: KRW 2 trillion for semiconductor production infrastructure, KRW 500 billion for physical AI demonstration zones, and KRW 4.7 trillion going directly to building domestic frontier models, KRW 7.2 trillion combined, with the rest unbroken out. So today I'll only discuss how that KRW 7.2 trillion is allocated, not the proportions of the total.
• Most surprising bit: Among the disclosed line items, the largest isn't buying compute, it's "training our own frontier models," effectively declaring that sovereign models are the entry ticket to the next round of national competition.
• Taiwan angle: Taiwan's strength is running other people's models fast and cheap; South Korea has chosen a more expensive, harder-to-reverse path. Worth discussing: the real barrier to building frontier models isn't budget, it's whether you can keep feeding in high-quality domestic-language corpora and retain training talent, and neither of those is any easier for Taiwan.
• Discussion points:
  1. When the state pays to train frontier models, who owns the IP and commercialization rights to the output?
  2. Is the KRW 500 billion for "physical AI demonstration zones" really about deploying robots, or is it showcase engineering?
  3. Should Taiwan follow suit and build its own, or double down on inference and vertical applications? That's a resource allocation choice.
• Script suggestion: "I especially want to be clear about the currency, because this kind of news is the easiest to get snowed by. The visible line items are three won figures: 2 trillion, 500 billion, 4.7 trillion, KRW 7.2 trillion total, with the rest unbroken out. So I'll set aside the 'how many US dollars total' question and just look at direction. And the direction is crystal clear: the biggest chunk goes to training frontier models in-house, not buying more cards. That's a strong policy statement, essentially saying model sovereignty sits at the same tier as energy and food. Whether Taiwan should take the same road: I think the real argument isn't about money, it's about corpora and people."

---

6. OpenAI trims its mega-round to ~$20B,butvaluationholdsat20B,\; but\; valuation\; holds\; at$1.25T

• Source: Mean CEO Startup Funding Report (citing Tech Funding News/CNBC; discovery source, aggregator-style, CNBC permalink pending) https://blog.mean.ceo/startup-funding-announcements-news-november-2026/
• Summary: OpenAI's raise, originally rumored at $30billion,hasreportedlyshrunktoaround30\; billion,\; has\; reportedly\; shrunk\; to\; around$20 billion, while the valuation stays at $1.25 trillion, led by a UAE sovereign fund. These figures currently come mainly from aggregator reporting citing other outlets and await first-hand confirmation. The timing is notable: Anthropic's IPO is approaching in the same week.
• Most surprising bit: The round shrinks but the valuation doesn't budge a cent, a rare combination in private markets.
• Taiwan angle: The AI-theme narrative driving Taiwanese equities has long rested on the premise that "model companies are still ramping aggressively." Once a "raised less than expected" signal appears, what gets hit isn't Taiwanese suppliers' current-quarter orders, it's the P/E multiples institutions are willing to assign. Sentiment moves first, fundamentals later.
• Discussion points:
  1. A smaller raise could mean weak demand, or it could mean the company is deliberately limiting dilution. The two readings lead to completely different places.
  2. With a sovereign fund as lead investor, does this plant a landmine for future geopolitical review?
  3. Pulling back on one side while sprinting toward an IPO on the other, is capital switching sides between two narratives?
• Script suggestion: "One reminder here: a smaller raise isn't necessarily bad news. It could mean the market can't absorb it, or it could mean the company decided 'I don't need to dilute this much.' But the valuation not dropping at all is the interesting part, because it means pricing power still sits with the seller. Personally I'd treat this as the most useful bubble thermometer this quarter: if over the next few months we start seeing deals that are both downsized and marked down in valuation, that's when the air is genuinely coming out. Until then, rather than listening to whoever's shouting 'bubble,' I'd rather watch how the terms change round by round. And again, that $1.25 trillion figure is secondhand reporting, so I'm holding back on it."

---

7. Anthropic prospectus reveals: Q3 revenue of $17B,purchasecommitmentsof17B,\; purchase\; commitments\; of$895B

• Source: CNBC (permalink for this report pending) https://www.cnbc.com/ai-artificial-intelligence/
• Summary: In the same S-1/A amendment, Anthropic disclosed Q3 revenue of $17billionalongsideroughly17\; billion\; alongside\; roughly$895 billion in purchase commitments, the vast majority of it compute. Separately, $3.5 billion of the US IPO shares are specially reserved for Japanese investors. These figures come from single-source reporting as far as I've seen, and still need line-by-line cross-checking against the original filing.
• Most surprising bit: A five-year-old company has committed to paying out more than fifty times a single quarter's revenue. That ratio alone is the most naked footnote to this arms race.
• Taiwan angle: The most overlooked aspect of these long-term contracts is "renegotiation risk." A purchase commitment isn't cash, it's a negotiable contract. The moment a model company decides to delay or renegotiate delivery cadence, the impact hits not the major chipmakers first, but the mid- and downstream suppliers who've already stocked materials and expanded lines on schedule. Receivables and booked capacity get squeezed together. When Taiwanese suppliers take these long orders, prepayment ratios and breach clauses matter far more than unit price.
• Discussion points:
  1. How should purchase commitments be understood on financial statements, and how far are they from debt?
  2. Reserving $3.5 billion for Japanese investors, capital considerations, or market positioning?
  3. If the revenue growth slope can't keep up with the amortization of commitments, which link gives way first?
• Script suggestion: "I know everyone's going to be shocked by that fifty-times ratio, but let me flip it around. A commitment isn't money already paid, it's a contract that can be renegotiated. So the real question isn't 'can they afford it,' it's 'if it gets renegotiated, who bleeds first.' The answer usually isn't the company at the very top of the chain, it's the mid-tier suppliers who've already expanded lines, stocked materials, and reserved capacity on schedule. Plenty of Taiwanese companies sit right in that segment. So my advice is: when you land one of these gorgeous-looking long-term orders, go look at the prepayment ratio, the breach clauses, and whether the delivery cadence can be adjusted. Those matter far more than the number on the quote."

---

Closing

What I think is most worth remembering today is that "the capability jump was only discovered after the fact" situation. More than any prediction, a retrospective confession is what should actually make you sit up straight. The other six stories, whether money is flowing in or drying up, whether it's governments stepping in or platforms taking things down, are all really about the same thing: the systems are moving faster than our instruments for observing them. Same time tomorrow, Mark's Tech Insights. See you then.