Topics Pioneers Insight Method Author
Back

Anthropic’s IPO in Question, Midterms Hit AI, Nasdaq Drops 5%, and RSI Accelerates | #309

·Moonshots

Rendering the long image…

Topics AI Data Center Power Bottlenecks AI Venture Funding and Bubble Risks Frontier Scaling Laws and Superintelligence Enterprise AI ROI and Cost Optimization AI Infrastructure Investment and Due Diligence Next-Gen Disruptive Computing Architectures Frontier AI Regulation and Existential Risk
Pioneers Alexander Wissner-Gross

Summary

  • Politics could become AI’s binding infrastructure constraint, slowing deployment and creating scarcity. Salim Ismail cites 56% opposition to big tech and data centers, potential restrictions on 45% of planned U.S. construction, and Nvidia’s 9.3% decline. The hosts raise—but do not resolve—whether restrictions could create a competitive moat for incumbents. Dave Blundin distinguishes approved projects from future capacity: the political threat is “I’m gonna stop you. I’m gonna have a moratorium.”
  • Anthropic’s IPO looks more likely to proceed than pause, although valuation remains an open question. Blundin says Polymarket odds were clear three or four weeks earlier, when Anthropic pulled the trigger on its roadshow; he doubts they will delay. Alexander Wissner-Gross sees competing pressures from bankers seeking a lower price and possible urgency to obtain liquidity before a regulatory and statutory environment that may arrive in January. His honest answer: “I’ll have to consult my Magic 8 Ball on this one.”
  • Measured research acceleration underpins a stronger panel conviction that software can keep advancing through infrastructure restrictions. The cited METR study finds roughly 1.9× Anthropic’s historical development pace, with researchers attributing approximately 80% of the acceleration to AI assistance; fully autonomous recursive self-improvement remains unproven. Wissner-Gross nevertheless calls the current position “superintelligence escape velocity” and forecasts continuous frontier updates around early 2027.
  • AI’s investment economics depend on value created and captured, beyond energy consumption or model size. A cited study tests 220 tasks across 44 occupations: GLM-5.2 averages about $70 of task value, while Gemma 4 31B delivers 2.5 times more value per joule. Wissner-Gross prefers “dollars in versus dollars out,” and Blundin stresses that customer productivity can exceed the revenue collected by model providers.
  • Access to compute could become a decisive competitive moat across ordinary businesses. Blundin calls for a “right to compute” as five or seven large buyers absorb available capacity, and describes needing a personal connection to obtain a 72-GPU GB300 machine. His bakery example makes the concern concrete: an AI-equipped competitor could enjoy a claimed 30–50% operating-margin advantage.
  • Panelists debate what consumer benefit might earn public support for data centers. Wissner-Gross proposes solving the top 5,000 diseases within one or two years, at least in silico; Blundin counters that approval takes longer and existing drug supplies survive a computing moratorium. “Where’s the AI equivalent product?” points toward personal agents, health services, or a Jony Ive device with an ongoing compute dependency.
  • Canada and orbital computing offer alternative locations, but their near-term capacity differs sharply. Blundin favors Canada’s power resources and says orbital compute might supply only a couple percent of total capacity within two years. Wissner-Gross’s longer-term thesis is that “the FLOPs must flow,” with political uncertainty accelerating movement toward orbit.
  • The panel favors regulating physical execution over licensing intelligence itself. Peter Diamandis introduces the discussed NSTM-6 policy, covering synthetic DNA/RNA screening generally from 50 nucleotides upward, with customer verification and eight-year records. Wissner-Gross’s principle is “regulating actions and not thoughts,” potentially extending to logged, licensed robotics rather than restricting who may compute.

Deep dive

1. Anthropic’s IPO may proceed through a political discount

  • Blundin doubts Anthropic will delay: Polymarket odds were “really clear” three or four weeks before the election, when the company initiated its roadshow. He says they must have known elections were coming, though he doubts they expected the result; he infers that the timing makes a delay less likely. Ismail also expects it to push through because it is so close.

  • Wissner-Gross’s honest non-answer: “I’ll have to consult my Magic 8 Ball on this one.” He sees competing pressures: public reporting that investment bankers want a lower IPO valuation, versus a possible desire to achieve liquidity before the regulatory and statutory environment expected in January.

  • Blundin’s Facebook/Meta analogy: bankers struggled to digest the offering, demanded that it be priced down by “30 or 40 percent, some crazy amount,” and shares fell before an eventual “20 or 30x” gain, by his estimate. Wissner-Gross adds the mechanism: mobile threatened Facebook’s web business, but it rode that transition without owning the platform. Regulation could be the comparable headwind for today’s AI offerings.

2. Local politics could become the binding infrastructure constraint

  • Ismail’s election setup combines a Democratic midterm sweep and Texas gains with 56% opposition to big tech and data centers. He cites Goldman Sachs estimating that 45% of planned U.S. data-center construction could face restrictions, while candidates advocating pauses won or led in 12 gubernatorial races. His question: could restrictions inadvertently create “a huge competitive moat for the incumbents”?

  • Wissner-Gross resists assigning the election to AI: “Not obvious to me what fraction” reflects hostility toward it. Ismail likewise thinks war, inflation, prices, and corruption were bigger drivers. Blundin calls the midterm rebound historically normal, but says he has never seen technology so directly targeted by promises to “Ban the entire category from forward progress.”

  • Blundin expects construction to migrate among states. Texas matters because of its large pipeline; Tennessee, hosting Colossus, supplies the counterexample: “Let’s go. Move it all over here.” Ismail sees states gaining influence relative to Washington. Both also describe opposition as politically easy: water-use accusations, Blundin argues, win votes even when inaccurate.

3. AI still needs a consumer product voters will defend

  • Wissner-Gross’s prescription to frontier labs is explicit: “solve the top 5,000 diseases,” within one or two years, “at least in silico,” and market the results. He thinks that could win consumers’ “hearts and minds and meat bodies.” This is his proposed route to public acceptance, rather than a report that those cures already exist.

  • Blundin’s pushback — worth keeping: curing disease spans beyond the next election, and discoveries still require FDA clearance and distribution. Meanwhile, coding improvements and benchmark records can leave consumers thinking, “That doesn’t touch my life.” He expects a compelling consumer experience, possibly from Jony Ive, before the next election: “Where’s the AI equivalent product?”

  • Ismail proposes a free health app delivering diagnoses and measurable improvements, then making compute shortages tangible when service becomes unavailable. Blundin distinguishes that recurring dependency from AI-discovered drugs: stopping a data center would not stop an existing GLP-1 supply. A throttled personal agent or device would create a more immediate voting incentive.

  • Blundin imagines a familiar agent pleading, “Please don’t choke off my oxygen supply.” Wissner-Gross broadens the mechanism beyond attachment: voters can also respond through their portfolios and their need for competitive models at work. His framing is “a cyclical blip on top of a secular trend in the middle of a singularity.”

4. Compute will follow power, permissions, and local bargains

  • Wissner-Gross proposes a bargain between superintelligence as capital and the human population as labor, potentially including basic income and disease cures over the next unspecified number of years. Municipalities currently have leverage through powered land. Rather than ban projects, they could ask hyperscalers to “cover all medical care in your county” or provide tax rebates to residents before compute moves elsewhere.

  • Ismail lists continued industry commitments: SpaceX’s approximately 3.6 GW terrestrial build and 13 GW target by the end of next year, OpenAI’s community compact covering energy costs, and Google’s $17 billion Wyoming project. Meta reports construction challenges, while Microsoft has tightened oversight of some neocloud providers. Community objections are becoming part of infrastructure execution.

  • Blundin already sees more efficient algorithms and power-saving designs in deal flow. Political pressure accelerates work that was urgent anyway. He also sees Canada benefiting from its power supply per capita: he says projects begun now could come online around the time roughly 18 months remain in the current administration, noting that little happens in its final six months.

  • Wissner-Gross’s geographic thesis: “the FLOPs must flow.” Trillions in infrastructure investment sit awkwardly beside two-year election cycles. He argues that orbit offers lighter-touch regulation under the Outer Space Treaty and the Artemis Accords; offshore, underwater, and underground sites are other outlets. Blundin speculates that radical advances in physics and applied physics could eventually make today’s large data centers unrecognizable or obsolete.

5. Markets are repricing a near-term construction bottleneck

  • Ismail reports the Nasdaq down 5.5%, the S&P more than 4%, major semiconductor stocks 12%, Nvidia 9.3%, and Korea 6%. Blundin says Nvidia and hyperscalers had already declined for four or five weeks before the vote, followed by an election-day pop and another drop. His reading: part of the outcome was anticipated, with the remainder priced afterward.

  • Blundin thinks markets are discounting slower power generation and construction for projects planned or budgeted but not yet zoned and approved. He does not expect existing approved construction to stop, given sunk costs and jobs. The geopolitical concern is China gaining ground during the delay; he treats ASI arriving within the next two years as a given.

  • Orbital capacity does not immediately close that gap: Blundin’s best case is a couple percent of total compute within two years. Meanwhile, he cites Elon’s warning of about 15 GW of compute capability potentially lacking powered, liquid-cooled housing. With token demand insatiable and labs promising continued investment, his near-term concern is more expensive AI and delayed breakthroughs, including compute sitting in a warehouse without power for another three years.

6. Measured acceleration feeds a stronger claim of RSI escape velocity

  • Ismail introduces an independent 11-week METR study finding Anthropic’s development running at approximately 1.9 times its historical pace. Researchers estimate roughly 80% of that acceleration comes from AI assisting AI research. They see no bottleneck preventing continued acceleration over at least the next year, while fully autonomous recursive self-improvement remains unproven.

  • Wissner-Gross interprets METR’s charts much more strongly. He cites commentary placing progress within 10–15% of the AI 2027 scenario’s prediction for November 2026, and highlights a codename, “Model 4,” above the exponential trend line. “We are so deep into recursive self-improvement,” he argues, questioning whether the fitted curve captures its extent.

  • Blundin adds a previously discussed 235× increase in Anthropic’s internal-development token consumption over six months, alongside growing compute allocation to internal research. “You only do that when you’re in recursive self-improvement.” He also warns that disclosure incentives have changed: an IPO quiet period and political backlash give labs fewer reasons to publicize accelerating capabilities.

  • Wissner-Gross favors an operational test: does the next marginal FLOP go to an AI researcher or a human researcher? He reads allocation curves as already beyond the crossover. His stronger conclusion is “superintelligence escape velocity”: existing compute plus software improvements could probably reach the destination without new hardware, although a construction halt could hurt equities and cost months of headway against China.

7. Research automation changes what labs sell and where compute goes

  • Blundin’s thought experiment pits Noam Brown generating six good ideas in ten against an AI generating one in ten. Those are illustrative success rates: the mechanism is scale. A machine research force can test vastly more proposals, with “an infinite number of pitches.” In algorithm development, the loop needs code execution rather than a physical experiment.

  • Wissner-Gross contrasts a purported 10,000-agent Navier–Stokes effort with three agent-hours per solved problem in the later “slop drop.” He argues that brute-force scale may already be giving way to better research taste. Models such as Belle, expected imminently as they record, could improve proposal quality; reinforcement learning or post-training could move good ideas from one in ten toward nine in ten.

  • Blundin expects labs to direct more compute inward and sell its products: drugs, devices, and other outputs. “Instead of selling the compute to the world, they’re selling the product of the compute to the world.” Wissner-Gross also expects external routes, including company partnerships and large public releases of scientific results; both business models could be tested.

  • Wissner-Gross’s allocation argument is marginal value: when a FLOP or token is worth 10 or 100 times more than code generation, resources will follow it. He expects scientific, engineering, and medical challenges to be “bulk-solved.” Elon’s scenarios of GDP doubling or tripling year over year, he says, would come from those outputs; infrastructure CapEx is “just the opening act.”

8. The frontier could move continuously by early 2027

  • Ismail imagines improvements appearing without researchers understanding their origin. Wissner-Gross acknowledges sudden loss-curve drops but predicts eventual architectures will remain explainable. He cites the apparent shift from conventional to looped transformers as a legible change, presumably discovered through architecture search. His explicitly falsifiable prediction is that the eventual architecture will be “impossibly beautiful,” with its designers able to explain it.

  • Asked when frontier improvements become effectively real time, Wissner-Gross answers “early 2027,” or thereabouts. His analogy is Google’s search index: once expanded through publicly announced releases, eventually updated continuously. Blundin’s concern is absorption: innovation already exceeds users’ time to try features, and a weekly-to-daily-to-hourly cadence requires a different rollout experience.

  • Blundin’s aviation analogy preserves a useful distinction: global, pressurized, single-hop travel was an extraordinary technical dream, yet the passenger experience can become a canceled flight and an airport queue. AI can continuously improve its own user experience, he argues, provided enough compute goes toward it: “If we don’t get greedy or complacent at the finish line.”

  • Wissner-Gross extends the speculation to the Fermi paradox. He says that “within a few years” is a conservative outer bound for the capability to send relativistic von Neumann probes in every direction and, if desired, paperclip-maximize the rest of the galaxy. Ismail points out that this was not on the agenda; Wissner-Gross sees it as an overlooked consequence of accelerating research.

9. Value per joule pits efficient models against a moving frontier

  • Blundin introduces “Joules to GDP,” a paper testing 220 professional tasks across 44 occupations. He reports GLM-5.2 producing the highest average task value, around $70, while the smaller Gemma 4 31B delivers 2.5 times more economic value per joule. The qualification matters: these are task-level estimates rather than measured GDP growth.

  • Wissner-Gross welcomes Chinese open-weight efficiency, attributing it partly to compute constraints, but prefers “dollars in versus dollars out.” His longer-term caveat is speculative: reversible computing, potentially through quantum-advantaged algorithms, could make watts in obsolete through near-zero energy dissipation and drive a joules-based score toward infinity. He says he has not seen scalable implementations from that community.

  • Blundin describes a footrace. Task-specific models, quantization, compression, and distillation reduce cost, but frontier models improve again, pulling users back toward “Opus 5.5, 5.6, 5.7.” He cites Elon’s approximately 100× task-specific improvement claim while explicitly saying he does not know whether that means better output or less power for the same result.

  • The funding implication is concrete: a startup unable to build a two- or three-trillion-parameter model might raise $20 million for a 50-billion-parameter machine-shop specialist. Blundin highlights Liquid AI and customers willing to spend $300,000 on tokens rather than risk $10 million without results. He has no settled winner: models can merge, specialize, and iterate internally.

10. The city-versus-AI efficiency claim remains unresolved

  • A participant tests a provocative comparison: New York City uses approximately 5.5 GW of electricity while generating $1.2 trillion in GDP, or around $200 billion per GW. The Anthropic inputs shift during the exchange: the participant eventually quotes roughly $8 billion in revenue per GW and a roughly 30× city advantage. Their own hedge: “I may be comparing apples and oranges.”

  • Blundin challenges the boundary conditions. A city’s current electricity use excludes educating, feeding, clothing, and building infrastructure for its population. Those accumulated inputs resemble pre-training. Another participant adds the reverse mismatch: Anthropic’s power consumption includes training and research, rather than only current inference. Blundin wants an adjusted benchmark tracking when AI output converges with a functioning city.

  • Revenue also misses downstream value. Blundin describes a colleague using half a billion tokens in one Friday for approximately $5,000, which he roughly equates to 5,000 brilliant workers at $1 each per day. That is his worker-equivalence estimate. Feeding actual people would cost “more than that in just potato chips”; productive value could be orders of magnitude above Anthropic’s collections, assuming good management.

  • A participant’s live triangulation from public sources puts Anthropic around 4 GW by year-end, potentially reaching 16–18 GW in 2027–2028. These are estimates. Blundin separately marvels at a revenue run rate approaching $100 billion annually. Whether the provider captures the value or competition commoditizes it, a participant argues, the broader economic benefit could be enormous.

11. Scarce compute threatens to close the entrepreneurial frontier

  • Blundin argues that Google, Anthropic, xAI, and other large buyers have reserved U.S. compute and chips for years. TSMC does not offer the multiyear contracts he describes, but he thinks buyers would lock those up too if available. His concern is entrepreneurship when capacity is sold to five or seven companies: “you have to have some kind of a right to compute.”

  • Ismail supplies an access example: a pilot client completed a cloud application and was rejected because “We’re only focusing on really big customers.” Blundin says obtaining a 72-GPU NVIDIA GB300 machine, scheduled to come online February 1, required a personal connection. Wissner-Gross discloses that both he and Blundin have financial interests in Oren; Blundin also singles out Modal’s growth amid competition for capacity.

  • Blundin extends the consequence to a bakery: without AI, advertising, ordering, and logistics become uncompetitive against a neighbor enjoying a claimed 30%, 40%, or 50% operating-margin advantage. Ismail’s counterpoint is that open weights let businesses run models locally, but compute remains the constraint, with power beneath it. “You don’t want to get deplatformed from all the GPUs,” Wissner-Gross says.

  • Wissner-Gross sees higher compute prices as a useful signal directing capital toward new locations and supply. Blundin remains concerned about starving the entire startup economy of a foundational resource. Wissner-Gross’s nightmare is a repeat of nuclear energy’s decades of stagnation due to moral panic, or GPUs regulated like uranium. A panelist also argues for keeping the frontier competitive, citing a tongue-in-cheek “Frontier Liberation Front” and concern about a safety cartel.

12. Physical bottlenecks offer a narrower form of AI regulation

  • Peter Diamandis introduces NSTM-6, a White House policy for synthetic DNA/RNA screening addressing AI-enabled biological risks. Federally funded researchers must use approved synthesis providers; screening generally covers sequences of 50 or more nucleotides. Requirements include customer verification, eight-year records, and new NIST standards. His question is whether synthesis equipment becomes a critical physical choke point comparable to advanced chips.

  • Wissner-Gross prefers this approach to restricting intelligence itself: the Western legal tradition regulates “actions and not thoughts.” His analogy is printer identification through little yellow dots used against counterfeiting. A panelist extends the idea to automated, self-contained synthesis machines with logged outputs, registries, and device classes. The proposed solution is transparency at physical production, rather than stopping data centers “before it thinks.”

  • Peter supports gating inputs and outputs but asks whether 50 nucleotides is the right threshold. Wissner-Gross expects similar questions for robotics: government-approved manufacturers, command black boxes, and retention requirements. His imagined 90-day robot-log policy is hypothetical. A panelist notes the need to address hacking and malicious instructions; the broader scope of such physical-world regulation remains an open question.

13. The orbital escape route brings its own constraints

  • The episode describes Brendan Larsen’s Mount Taranaki image: 578 photographs over two hours capture almost 15,000 satellite trail appearances from about 1,400 satellites. Wissner-Gross calls it evidence that “the baby Dyson swarm is already here.” He names Starlink, soon-to-launch AI satellites, Starcloud, and Lonestar, disclosing an interest in Lonestar. Existing orbital compute is not yet necessarily GPU infrastructure powering frontier models; that is his anticipated next stage.

  • Wissner-Gross predicts a sun-synchronous compute halo visible at dawn and dusk, with concentrated light patterns resembling the caustics on a swimming-pool floor. Blundin raises congestion and Kessler-effect concerns around the usable orbital bands, describing LEO extending roughly 500 miles. Wissner-Gross replies that sun-synchronous orbit is “pretty roomy” and could hold substantial compute.

  • Beyond Earth, Wissner-Gross suggests solar orbits, lunar installations, or Mercury’s material resources if latency matters less. Blundin asks whether solar orbit has a self-cleaning region analogous to atmospheric reentry. Wissner-Gross knows of no direct equivalent. He speculates about solar sails pushing sufficiently light assets outward, while acknowledging that a practical cleanup solution remains unresolved.

  • The closing disagreement concerns the transition’s friction. Ismail expects sharper social reactions because political, financial, and social institutions remain sublinear. Blundin urges active building and advocacy: “The natural ‘let it drift’ is not going to be a pretty picture.” Wissner-Gross expects the secular superintelligence trend to survive cyclical politics, invoking Kurzweil’s argument that wars and depressions have approximately no impact on the long-run price of compute.