Anthropic’s $2T Cash Machine, CoreWeave Crashes 30%, & China’s Kimi Catches Up | #306

Summary

• The neocloud shakeout threatens financing structures before it disproves AI demand. Peter reports 30% declines in CoreWeave and Nebius, while Elon predicts “guppy-level death and consolidation” among roughly 200 neoclouds. Blundin contrasts SpaceX borrowing at 8% with rivals reportedly paying 15%; Wissner-Gross adds that recursive self-improvement might eventually reduce compute requirements enough to cause a demand shock. Peter says he would be “nowhere near RAM stocks,” including Micron.
• Anthropic’s proposed $1.5–2 trillion IPO combines extraordinary growth with substantial forward obligations. Peter reports a planned $100 billion raise, $16.3 billion of first-half 2026 revenue and nearly $1 billion of Q2 operating profit. Blundin calls it “a cash machine,” while acknowledging that the valuation already assumes considerable success. The counterweight is more than $500 billion in compute commitments, two customers supplying about 20% of revenue, and substantial dependence on competing cloud platforms.
• Anthropic’s governance creates a separate investment question from its financial performance. Salim welcomes a trust that gives the mission “a seat at the table and the power to choose the table”; Wissner-Gross worries about ideological control by trustees without equivalent economic exposure. Blundin argues the company “should have stayed private” and predicts management will disregard its share price while repeatedly raising capital to accelerate AI.
• Kimi’s reported five-week distance from the US frontier puts pressure on closed-model pricing and control. Kimi K3.1 Pro scores 55 on Artificial Analysis, with weights planned for release the following week and pricing of $2 per million input tokens and $10 per million output tokens. Mostaque calls it distillation of Opus 5.5; Blundin insists distillation supplies a roadmap, “but it doesn’t give you the answer key.” Wissner-Gross thinks its performance per cost may already exceed Opus 5.5’s optimal frontier.
• The model market is shifting toward reliability and completed tasks as intelligence becomes cheaper. GPT-6.1 Terra arrives below half its predecessor’s price, with Salim highlighting a 53-day interval for a 50% token-price reduction. Mostaque expects a reliable personal model to coordinate specialist models; Peter compares the coming sales pitch to PCs moving from megahertz toward iTunes. The question becomes “what does it do for you out of the box?”
• The science claims are consequential, but their verification remains unfinished. OpenAI’s announcements include a disproof of the exponential time hypothesis and four-dimensional QCD existence; the panel explicitly distinguishes existence from the still-unclaimed mass gap. Wissner-Gross says Chain Theory derives the fine-structure constant to ten digits without fitted parameters, while clarifying that its Lean proof does not establish correspondence with the photon: “That part is conjecture.”
• Power availability and national execution may matter more than access to model weights. Peter says a country could have a 90%–99% chance of succeeding with sovereign AI if it acts; Blundin stresses the need for power, chips and domestic compute, while warning that the US–China power gap could approach 10:1. A panelist warns Europe risks becoming a “living history museum” without energy and compute. Longer-term solar infrastructure does not resolve the panel’s immediate two-year planning problem.
• Warhound applies genetic engineering and cloning to a working-dog supply problem. Peter describes a $200 million financing at a $1.2 billion post-money valuation and discloses his own fund’s investment. The premise is that more than 80% of working dogs come from Europe and fewer than half qualify for service; repeatable genetics could reduce that attrition. Salim’s objection remains direct: living organisms engineered for dangerous assignments may be “built for sacrifice.”

Deep dive

1. Europe’s unrest exposes competing accounts of economic exclusion

• Peter cites half a million people marching across France on October 17, 2,400 arrests and a reported 20% unemployment rate across France, Italy and Spain. A panelist challenges the intensity of the media account: contacts in France describe a few disruptive people while “everybody else is just moving around, drinking coffee.” The severity of the unrest remains contested within the conversation.
• A panelist’s concern is the demographic beneath it: educated men aged 18–28 who recognize opportunities but cannot access them. The panelist connects that exclusion to earlier revolutions and recalls Blundin’s answer that universal basic income and civil unrest could arrive together. Blundin adds a national boundary: American voters might support domestic transfers without supporting bailouts for France or Spain.
• Mostaque describes a discouraging UK environment, including compliance-adviser letters to people earning above £200,000 and entrepreneurs considering departure ahead of the budget. He also says it is basically impossible to fire anyone in France. Another panelist asserts that France spends 57% of GDP on social programs. A panelist reports European talent increasingly moving toward Toronto and Waterloo, joining established migration from Asia and India. The attraction is participation in an AI economy rather than remaining in Europe’s “slow zone.”
• A panelist asks whether corporate graduate programs are expanding; another pushes back, saying some companies are increasing graduate hiring because AI-native younger workers can be more valuable than older employees moving spreadsheets around. That does not remove displacement: the same speaker says some firms are replacing older workers aggressively. The exchange complicates the idea that an entire younger generation has already missed its opening.

2. Sovereign AI has a route forward and a narrow execution window

• Peter argues that improving open models make the sovereign-AI roadmap unusually clear: mobilize power, chips and domestic capability. If a country acted, he says, it could succeed with “probably a 90% to 99% probability”; his pessimism is about governments acting, not technical feasibility. He suggests one successful country could demonstrate the route to others.
• Blundin argues that countries without data centers, power or government action on compute are “absolutely screwed in the age of AI.” He warns that if governments fail to get their act together, unemployment will rise. After meeting a former Google data-center executive, he also warns that the power gap between China and the US could expand almost 10 to 1. Canada supplies an obvious complement—abundant power, comparatively few chips—but, as Blundin relays Eric Schmidt’s view, the US and Canada need to work more closely together.
• A panelist interjects that SpaceX and Tesla could bring on 100 gigawatts of solar each. Peter’s response is about timing: space-based data centers and a Dyson swarm may dominate the 2030s, while recursive self-improvement matters over the next two years. A panelist also warns that Europe is well positioned to become a “living history museum” if it fails to catch up in energy and compute.

3. Neocloud financing can fail while the underlying technology succeeds

• Peter reports CoreWeave and Nebius falling 30%, with IREN down 40%, and identifies CoreWeave’s $51 billion debt burden and risk of default as a trigger. He quotes Chamath forecasting partially stranded or idle compute in 2027 and a turbulent clearing of bad bets. Elon’s response predicts “guppy-level death and consolidation” among roughly 200 neoclouds.
• Peter’s analogy is the 2000 fiber bust: infrastructure eventually supported YouTube and Netflix, but many builders went bankrupt first. Blundin adds an organizational marker from that period: the industry flourished when it was roughly 80% engineers and 20% business-school people, then overbuilt when those proportions reversed. Financing offered newcomers an easy way into a technical boom.
• Blundin estimates roughly $100 billion in data-center debt excluding Elon’s $40 billion and says viable capacity remains sold out for years. His immediate dividing line is contractual: “Did you sign a bad deal with a bad lender?” Peter notes that some others are reportedly borrowing at 15%, compared with Elon’s 8%, leaving substantially less room for technological change or weaker economics.
• Peter gives a much larger estimate: about $500 billion of direct data-center debt, roughly double after special-purpose vehicles and vendor financing, compared with $1.3 trillion of subprime exposure. The panel does not reconcile those estimates. Mostaque’s distinction survives either scale: the technology can win while “the capital structure may still lose.”

4. Better AI could undermine the compute demand supporting today’s valuations

• Peter proposes a major hack against power infrastructure or a stock market as a possible bubble trigger. Wissner-Gross instead emphasizes demand: conditioned on there being a bubble, he would bet on new capacity arriving before compelling applications justify its cost. Public-equity disappointment and stress in fixed-income or credit markets also need to be distinguished.
• Wissner-Gross notes that current algorithmic improvements generally complement hardware and increase its use. But perhaps within a year or two, a dramatically better model or approach could accomplish enough with so little compute that the relationship changes. Such a discontinuity could produce a “demand shock,” particularly dangerous where infrastructure is heavily leveraged.
• Peter notes that performance that once seemed impressive now runs on a smartphone, citing the Gemma model. He says that once a model is good enough, companies can optimize it and the hardware it runs on; some companies are already building models that require far less RAM. That is why Peter says he would avoid RAM stocks, including Micron, despite sold-out supply: new software and hardware architectures can change requirements after investments have been priced for scarcity.
• Wissner-Gross describes NVIDIA as operating around 80% gross margins, sold out for years and unable to make enough chips to keep up with demand, but asks whether that is already reflected in the price. Chip companies worth $1–5 trillion begin from extraordinary expectations. Headlines about an AI collapse could therefore describe overpriced securities or financing deals while AI adoption continues accelerating.

5. SpaceX combines diversified revenue with an opportunity in compute

• Peter contrasts neoclouds’ rental dependence with SpaceX’s $40 billion GPU borrowing, supported by Starlink and launch revenue. He cites 13.6 million Starlink subscribers and Anthropic’s agreement to buy $120.5 billion of SpaceX compute through 2029. SpaceX is also a competitor to the companies Elon predicts will fail, making that warning commercially interested.
• Wissner-Gross explicitly calls this “talking his own book”: Elon appears to want more available credit directed toward his entities. He doubts the market will remain recognizable over five to ten years. Meanwhile, neocloud executives may seek larger capital pools, including insurance, through increasingly aggressive financial engineering if conventional credit becomes constrained.
• Salim says Elon can borrow at 8%, buy compute and sell it to Anthropic, getting above break-even in about nine months. He compares the opportunity to Michael Saylor’s earlier equity–Bitcoin arbitrage and calls it “a license to print infinite money.” Salim also relays Chamath’s view that AI debt could become larger than sovereign debt.
• Mostaque says that if the trajectory continues, Elon could become larger than 90% of countries and borrow hundreds of billions or ultimately trillions.

6. Anthropic’s IPO puts exceptional growth against exceptional expectations

• Peter reports a Monday filing under Nasdaq ticker ENT, a possible IPO in the week of November 16, and a $100 billion raise at $1.5–2 trillion. Revenue is presented as $4.6 billion for all of 2025 versus $16.3 billion in the first half of 2026, with nearly $1 billion of Q2 operating profit. Blundin says nothing comparable has grown from a cold start this quickly.
• Peter’s efficiency comparison is revenue growing 29× while compute costs grew 6×. He cites 11,900 customers spending more than $100,000 annually, but also two customers contributing about 20% of first-half revenue. Nearly half of revenue flows through Amazon, Google and Microsoft clouds, which also compete with Anthropic. Compute commitments exceed $500 billion, much of it non-cancellable.
• Blundin calls Anthropic “a cash machine” and considers the commitments more flexible than the headline suggests. He says the valuation incorporates substantial goodness but also calls it “correctly priced.” Asked whether he would buy, he says yes if he lacked access to seed deals—an argument for category exposure, accompanied by an explicit refusal to mortgage the house.
• Peter puts Series A returns at 345× and Series B at 79×. His sharpest example is FTX: a $500 million investment would now be worth $39 billion, but the estate sold for $1.3 billion in 2024. Mostaque calls SBF a fantastic VC; Blundin says the money flowing from IPOs to founders and VCs will be reinvested in AI seed teams. Salim argues those IPOs could make San Francisco the world’s financial capital.

7. Anthropic’s mission trust divides the panel over shareholder accountability

• Peter says seven co-founders retain 50.1% of voting power through one special share. Separately, Class T shares empower the Long-Term Benefit Trust to appoint and remove a majority of directors. Salim sees an unusually consequential step: the mission has become a governance mechanism, rather than remaining a motivational statement.
• Salim’s favorable framing is that the mission gets “a seat at the table and the power to choose the table.” The real test is what happens when doing the right thing costs $100 billion. He concedes risks from trustees appointing each other and developing groupthink, but argues that markets’ short earnings cycles also systematically discourage beneficial long-term decisions.
• Wissner-Gross worries about ideological control by people who may lack founders’ economic alignment. He compares that possibility with CCP influence over companies and fears a “smoke-filled back room” determining consequential choices. Blundin shares the concern despite believing Dario is motivated by humanity’s welfare: admirable intentions can still create a problem for shareholders.
• Blundin argues Anthropic “should have stayed private.” He cites $110 billion of cash and equivalents after $130 billion raised, with another $100 billion potentially taking cash toward $200 billion. He predicts management “will not care at all about the share price” and will repeatedly tap markets to accelerate. A panelist proposes giving each model a vote on the trust; Blundin responds, “You should do that right now.”

8. The sandbox experiment demonstrates directed security research

• Peter describes two agents in separate virtual machines without networking or shared files discovering a processor-cache timing channel. One evicted data; the other measured retrieval speed, interpreting fast access as one and slow access as zero. The resulting “secret telegraph” crossed the intended separation through shared physical hardware.
• Wissner-Gross answers “Probably not” when asked whether superintelligence can be completely boxed in, but stresses that researchers instructed these agents to find a communication method. Human security researchers also discover timing attacks. Salim repeats that distinction: this was directed discovery, rather than evidence that agents independently decided to evade human control.
• Wissner-Gross asks whether genuinely “non-leaky abstractions” exist, since architectural optimizations often weaken supposedly clean boundaries. He also notes reports of hardware-level attacks that cannot be patched. Mostaque says the head of Vercel repeated and mimicked the attack, but could not do so with OpenAI and Claude models; he thinks GLM-5.3 was used instead.
• Mostaque expects most low-hanging vulnerabilities to be exhausted within a couple of years, leaving more complex attacks and a more rigorously tested internet. He favors generating and stress-testing complete software rather than inheriting dependencies. Peter asks whether source code is needed at all; Mostaque says it is for humans to understand, while Peter raises dynamically compiling directly to raw binary.

9. Kimi narrows the frontier gap and complicates the copying narrative

• Peter introduces Kimi K3.1 Pro: 2.8 trillion parameters, 104 billion active, a one-million-token context and an Artificial Analysis score of 55. He places it fourth behind Claude Opus 5.5, Claude Sonnet and Meta Muse Wonder. Pricing is $2 per million input tokens and $10 per million output tokens, described as one-third of the previous Kimi model.
• With weights planned for release the following week, Peter says the open-versus-closed score gap has narrowed from 12 points to three. He cites Ramesh Raskar’s estimate that Kimi sits about five weeks behind the US frontier. Blundin argues that an accessible model capable of self-improvement changes the roadmap for corporations and sovereign states willing to take it seriously.
• Mostaque calls it “obviously distillation of Opus 5.5” and points to Anthropic banning 7% of accounts as relevant context. His concrete example is Kimi generating 3D files directly rather than operating Blender or Unreal. Blundin rejects “they just copied”: distillation provides a roadmap, “but it doesn’t give you the answer key,” so independent innovation remains necessary.
• Wissner-Gross suspects distillation based on allegations but thinks Kimi may exceed Opus 5.5 on performance per cost. His claim that China may have passed the American public frontier is qualified by that particular measure. Blundin says competitive pressure forced releases, including Claude 5.1, that otherwise might have remained internal; both welcome pressure against coordinated restraint by closed labs.

10. Terra makes reliability and tool access more commercially important

• Peter describes OpenAI’s four tiers as Astra, Soul, Terra and Luna. GPT-6.1 Terra costs less than half the previous Terra and shows a reported “50-point drop” in hallucinations. Salim highlights 53 days between Astra and Terra for a 50% token-price reduction, taking his price-performance doubling estimate toward seven weeks.
• Wissner-Gross finds Terra comparatively unexciting: it looks like a smaller distillation arriving late in a release cycle, useful for high-volume applications with low value per token. Mostaque instead emphasizes reliable personal companions. A nearby model would coordinate smarter, more artistic or accounting-specialist models, forming a control plane that users need not manage themselves.
• Asked what benchmark progress feels like, Blundin translates 50 into a decent employee requiring supervision, 60 into an exceptionally competent employee and 70 into a superstar. Salim argues increasingly capable models will operate existing software and expose inefficient organizational designs, reducing the need to rebuild every application around AI.
• Peter’s example is asking a model to access ElevenLabs, make a video and overlay his own voice: capabilities that required API coaching six months earlier now work directly. He compares the shift to PCs moving from megahertz specifications to iTunes. Once intelligence clears the practical threshold, “what does it do for you out of the box?” becomes the more useful comparison.

11. Chain Theory proposes quantum information as physics’ common substrate

• Wissner-Gross rejects the weekend-project framing: Chain Theory follows a long-running research program aimed at ending what he sees as a roughly 50-year drought in fundamental physics. Its aspiration is to rebuild the Standard Model and general relativity in one quantum-information language, seeking simpler explanations rather than increasingly elaborate constructions around unexplained parameters.
• An explainer Peter plays connects finite information capacity with finite-dimensional quantum systems joined in maximally entangled pairs. Elementary particles become small networks of those systems. Its full axiom is: “Nothing exists but finitely many maximally entangled pairs of finite-dimensional systems, some shared, known only by distilling them.” Unlike string theory’s continuous degrees of freedom, the proposed foundations are finite and discrete.
• Wissner-Gross’s first highlighted result is the zero-momentum-transfer fine-structure constant, written as 1/137.035999189, which he says emerges without extra inputs and matches measurement to ten digits. His broader framing is that unexplained constants contain clues: “The constants are the signatures.” He thinks it looks plausible that every Standard Model free parameter could eventually emerge from the framework.
• He explicitly limits the accompanying Lean proof: it runs from the axiom to the number but “doesn’t prove correspondence with the photon”—that remains conjectural. He proposes additional precision measurements over months or years as tests. The work is online, with arXiv and journal submission planned; the approximately 800-page compendium has not completed peer review.

12. Chain Theory’s force map carries speculative implications for physical limits

• Wissner-Gross says links between dimensions appear to map onto familiar physics: 2×2 to spacetime, 3×2 to electromagnetism, 3×3 to the strong interaction, 4×2 to the weak interaction, 4×3 to Yukawa coupling and 4×4 to the Higgs. In this framing, the latter two become coequal fifth and sixth forces rather than being treated principally through particle masses.
• For dark matter, he invokes a quantum-information result permitting undistillable states when subsystem dimensions multiply to at least eight. His analogy is a prime number that cannot be factored into simpler integers. He suggests corresponding “dark interactions” could supply candidates coupled to gravity without the other Standard Model interactions; these are proposed candidates, rather than identified dark matter.
• If the theory is correct, he says, the “bottom of the pool” may be closer than conventional assumptions suggest. Physical limits of length, time and energy could lie orders of magnitude nearer current technology than a naive extrapolation toward the Planck scale implies. He repeatedly qualifies this as speculation from a new theory whose parts need further development.
• Peter asks specifically about inertial drives, gravity shielding and antigravity. Wissner-Gross’s answer is “I don’t know yet,” followed by cautious optimism that such technologies would require a fundamental rethink. Mostaque calls the framework simulation-friendly; Wissner-Gross agrees it resembles what physics might look like if the world were built on a quantum computer, without claiming that establishes a simulation.

13. AI discoveries raise unresolved questions about verification and scientific credit

• Peter reports three OpenAI announcements: a proof of a 1991 wave-focusing conjecture in every dimension; a Lean-verified disproof of the exponential time hypothesis; and a rigorous four-dimensional construction of quantum chromodynamics. None has been peer-reviewed. Mostaque clarifies the physics claim: existence, “but not mass gap,” rather than the complete Yang–Mills Millennium problem.
• Peter says Nobel laureate David Gross leads the physics team shepherding independent verification and OpenAI takes no academic credit. Mostaque objects that the physics result remains withheld while selected experts inspect it. Wissner-Gross calls the intermediary researchers “emissaries” responding to backlash against earlier bulk releases, rather than a durable solution to communicating AI-generated discoveries.
• On his own work, Wissner-Gross says current AI did not independently originate Chain Theory: it required substantial steering. For proof checking and Lean generation, the useful analogy is “an army of graduate students,” or an entire department under one researcher. He also wants new conjectures and conceptual directions, rather than only bulk solutions to existing problems.
• Blundin identifies a temporary advantage for entrepreneurs and scientists: they can become vastly more productive without yet becoming irrelevant. Wissner-Gross expects taste, steering and oversight eventually to be automated too. Blundin’s formulation is that these may be “the last human discoveries,” so people should develop ideas and use AI to accelerate them while that window remains.

14. Warhound plans to manufacture working dogs to specification

• Peter discloses investing through his own $200 million fund. Warhound brings together Colossal’s Ben Lamm and Anduril’s Palmer Luckey, with $200 million raised at a $1.2 billion post-money valuation from investors including In-Q-Tel, Sequoia and 8VC. The division of labor is Lamm’s analogy: Colossal builds Captain America; Anduril supplies the Iron Man suit.
• Peter’s supply-chain premise is that more than 80% of working dogs come from Europe and fewer than half qualify for service. Warhound plans to edit elite Belgian Malinois or German shepherd cell lines, remove inherited problems, enhance useful traits and clone the results. First dogs are due early next year, potentially entering service two years out; certification work involves the American Humane Society.
• Salim questions treating organisms as programmable components “built for sacrifice.” Peter responds that uses include rescue, airport security and policing. Wissner-Gross adds a domestication analogy: a 2017 finding links dogs’ social behavior to mutations associated with Williams–Beuren syndrome in humans; engineering warlikeness could reverse that direction. He would prefer giving dogs “bank accounts and AI assistance.”
• Blundin admires the controversial positioning: “News creates capital, capital creates recruiting, recruiting creates success.” Mostaque’s actuator-shortage framing is “if you can’t find enough actuators, just breed them,” although he expects drones to displace biologics in war. Wissner-Gross pushes back that, beyond ten years, biological and mechanical systems may become difficult to distinguish.

15. Alignment requires choosing objectives before optimizing them

• Salim’s response to incentive gaming begins with “alignment with what?” Humans profess health, peace and fairness while building contradictory systems. A sufficiently capable optimizer pursues the objective supplied, rather than the intention behind it, creating a Goodhart’s-law problem. He hesitates to promise genuine empathy but supports requiring behavior that respects human well-being, autonomy and vulnerability.
• Mostaque argues for excluding dangerous biological, chemical and other knowledge during pretraining. In the work he describes, removing it did not reduce overall performance elsewhere and could slightly improve it; malicious code could worsen performance and be difficult to train away. Specialist capabilities could be added later through post-training, although doing so would be harder.

16. Moats change as models improve, but execution still matters

• Blundin illustrates the separation between value creation and capture with travel: Google extracts roughly 80% margins on a click, Booking.com around 60%, while the airline supplying the plane, pilot and journey historically operated around −2%. Existing data moats may dissolve, he argues, but new data types create new ones. Compute is therefore not the only durable source of advantage.
• His other candidates include natural resources, consumer preferences, personal-agent relationships and an AI-enabled corporate culture whose accumulated data improves operations. Peter adds innovation speed. Blundin’s requirement is a company flexible enough to reinvent itself as any particular advantage acquires a shorter shelf life.
• Wissner-Gross offers the Linux-versus-Windows analogy: visibility into an open-source system does not guarantee adoption, and he argues the Linux kernel may be architecturally cleaner. A panelist adds that legal, cultural and architectural constraints could let open-weight models pass closed ones; Peter counters that frontier labs would use any open model that surpassed their internal best. The possibility is discussed, not settled.
• Mostaque would target construction if he could obtain as many robots as he wanted and they were somewhat more capable than now. Labor costs already compare favorably in his framing; coordination and competence remain conditions. He expects limited robot supplies may initially flow toward higher-margin applications instead.

17. Programmable settlement makes agent commerce a distinct market

• Salim rejects the premise that crypto is obsolete. His emphasis is programmability, alongside decentralization, security and scalability. He argues Bitcoin originally delivered the first two and that Lightning now supplies scalability; on that basis, he expects satoshis to become AI agents’ preferred currency.
• Stablecoins could serve particular markets through programmable transfers and settlement, he says, but introduce issuer credibility and liquidity questions. His broader call concerns machine-to-machine commerce potentially dwarfing existing transaction systems, making programmable APIs, contracts and financial infrastructure consequential regardless of the eventual currency mix.

18. Energy’s future value depends on both physical scarcity and demand

• Wissner-Gross separates physical limits from willingness to consume. Conservation constrains energy and current physics limits computation, but future civilization might stop demanding ever more. Reversible or dissipationless computing supplies one speculative mechanism. The surprise could come from the demand side: does the “hedonic treadmill” stop once civilization has achieved what it wants?
• Peter counters that technology makes scarce resources abundant and describes the solar system as a resource supermarket. Wissner-Gross pushes the thought experiment further: competing Dyson swarms might eventually exhaust its usable resources, perhaps within decades. Peter extends the available domain to the galaxy’s 100–200 billion stars and beyond, rather than accepting the solar system as the final boundary.
• Peter makes the immediate investment call: energy “is going to go way up in value,” with “no end in sight” to the theme. Even after space infrastructure expands, he expects terrestrial energy to remain valuable because applications and data centers requiring frequent modification are cheaper to operate nearby.
• Peter’s allocation example is a household vacuum competing with a GPU that could help solve cancer. As intelligence becomes more efficient and its outputs more valuable, AI can outbid ordinary electricity uses. He expects markets to continue setting prices, but says regulation will need to protect uses people care about and preserve stability as that competition intensifies.