Summary

Anthropic is gaining private-market momentum even while OpenAI remains the larger consumer franchise. Secondary demand was cited at $2 billion for Anthropic versus $600 million for OpenAI; Anthropic was being priced near $600 billion, up from $380 billion, while OpenAI traded roughly 10% below its $852 billion raise. The panel framed the contest as installed base versus intelligence: OpenAI has 900 million users and enormous cash, but Anthropic may be inventing “the new playbook of the future.”
xAI is attempting a foundational rebuild only months before a predicted $2 trillion IPO. Eight founding engineers, including three co-founders, have left; SpaceX personnel are filling leadership gaps as Elon Musk concedes xAI “was not built right the first time around.” Colossus 2 reportedly runs roughly 700,000 GP200s and GP300s representing $18 billion of hardware. Alex Wissner-Gross argued that the disclosed 10-trillion-parameter ceiling shows “the parameter scaling race seems to be over”; the discussion also suggested the largest model could serve as a teacher model for distillation.
Musk’s $100 billion lawsuit could determine OpenAI’s leadership, corporate form, and IPO path. Musk alleges fraud and breach of contract, seeks Sam Altman and Greg Brockman’s removal and a nonprofit reversion, with jury selection scheduled for April 27 in Oakland. Salim Ismail called it “a governance war, disguised as legal war”; the panel’s settlement scenarios ranged from giving Musk IPO equity to Peter Diamandis’s prediction that Altman steps down while OpenAI remains for-profit.
The AI financing boom is becoming a market-wide liquidity event, not merely a venture-capital story. The panel cited $242 billion of global AI investment in Q1 2026, 64% concentrated in OpenAI, Anthropic, xAI, and Waymo, with the run rate approaching “$3 billion a day.” Dave Blundin’s warning was that institutions cannot fund $50 billion-$100 billion allocations from idle cash: large public holdings may have to be sold, even while record venture dry powder keeps startup financing abundant.
The panel sees white-collar displacement as near-certain but remains divided over whether it produces a net jobs bust. At an MIT panel, Google DeepMind’s Peter Denenberg reportedly put the odds of replacing a randomly selected white-collar job within two years—at 10 times productivity—near 99%, before Liquid AI’s Alexander Ermini observed that this estimate used today’s technology. Against Marc Andreessen’s claim that job-loss narratives are “all fake,” the Moonshots group expects violent category churn, smaller firms, and perhaps a net boom in one-person AI conglomerates.
Managed agents and AI-native organizational design are the revenue bridge behind the trillion-dollar forecasts. Claude managed agents aim to execute complex workflows continuously, shifting enterprise spending “from software licensing to outcomes”; Wissner-Gross sees the race as Anthropic “becoming OpenClaw faster” than rival labs. The operating benchmark is also escalating: being an AI company is no longer sufficient—investors increasingly want recursively self-improving systems, revenue traction, and measurable machine leverage per employee.
AI is spreading from software into biotech and robotics through teams, data, and physical infrastructure. Anthropic reportedly paid $400 million for 10-person, eight-month-old Coefficient Bio, while Eli Lilly’s $2.75 billion Insilico Medicine deal included $115 million upfront and milestone-heavy economics. China leads available humanoid hardware—Agibot reportedly shipped 10,000 units—while the US retains an advantage in VLA foundation models; the bottleneck is now motors, components, factories, and deployment capacity.
Quantum computing is a solvable Bitcoin upgrade problem, but AI-driven irrelevance may be the deeper thesis risk. Google’s cited RSA timeline moved from 2035 to 2029, prompting Brian Armstrong’s proposed $150 million BIP-360 coalition; Michael Saylor argued that “the upgrade will come before the threat does” while buying 88,000 Bitcoin for roughly $7.25 billion. Dave Blundin agreed that quantum was manageable, while Wissner-Gross questioned whether fast-moving AI agents will use Bitcoin at all rather than invent currencies optimized around compute, energy, and machine-speed settlement.

Deep dive

1. xAI is rebuilding under an IPO clock

Peter Diamandis framed the setup as unusually fraught: eight founding engineers, including three co-founders, have departed, SpaceX personnel are filling leadership gaps, and a summer IPO at a predicted $2 trillion valuation remains in view. Musk’s own diagnosis was blunt: xAI “was not built right the first time around.”
Dave Blundin distinguished Musk’s proven strength in physical scale from the finickiness of model training. Colossus—built in record time with a reported million GPUs—fits the Tesla and SpaceX playbook; a software defect can instead waste an entire run, as illustrated by OpenAI’s rumored $500 million O3 training failure in 2024.
Wissner-Gross’s speculative diagnosis was that earlier Grok models “smell like they’re benchmarked”: strong on selected tests, but potentially optimized through hand-curated data rather than general reasoning. With Meta also struggling to convert compute into frontier performance, catching up requires more than replacing personnel or building the largest cluster.
Wissner-Gross’s organizational reading was that “the org chart is now part of the product stack.” Musk can immediately see a rocket explode or a Cybertruck window crack, but leaders can be misled more easily about training quality; AI teams have more room to “blow smoke up your ass” than manufacturing teams.

2. Brute-force model scaling has reached a ceiling

xAI’s disclosed program spans Imagine version 2, two 1-trillion-parameter variants, two 1.5-trillion variants, a 6-trillion frontier LLM, and a 10-trillion model. Colossus 2 was described as roughly 700,000 GP200s and GP300s, with an estimated $18 billion invested in hardware.
Wissner-Gross welcomed the transparency because other frontier labs largely stopped publishing parameter counts. Yet a 10-trillion upper bound, rather than hundreds of trillions, implies that “the parameter scaling race seems to be over,” much like clock-speed scaling plateaued in conventional computing.
The likely purpose of the 10-trillion model is not direct deployment but teaching: train the largest feasible model, then distill its capabilities into cheaper systems. Reasoning training and distillation—not raw parameter accumulation—now carry more of the improvement burden.
xAI has also abandoned the low end that Google serves through small Gemma models. Its portfolio reflects Musk’s characteristic preference for brute-force scale, while video generation remains separate from the general reasoning stack rather than becoming a first-class modality.

3. Musk versus Altman is a governance trial with IPO consequences

Musk’s $100 billion action alleges fraud and breach of contract against OpenAI, Altman, and Brockman. He is also seeking the executives’ removal and a return to nonprofit status; jury selection is scheduled to begin April 27 in Oakland federal court, with Musk, Altman, Brockman, and Satya Nadella expected to testify.
Ismail rejected a conventional startup framing: “This is a governance war, disguised as legal war.” The underlying question is who steers systems with “quasi-civilizational impact,” making the dispute more geopolitical than a normal founder or shareholder conflict.
Discovery reportedly surfaced a 2017 Brockman diary entry saying the nonprofit commitment was “a lie,” which Peter said helped the case proceed. Yet a New Yorker investigation also reported that Musk sought majority control of a for-profit in 2017, complicating his presentation as defender of the original nonprofit mission.
Peter Diamandis offered IPO equity for Musk as one possible settlement bridge and predicted that the parties settle, Altman steps down, and OpenAI remains for-profit—because Musk may care less about $100 billion than delivering “the bullet to Sam and Greg.”

4. OpenAI’s nonprofit conversion could set a sweeping precedent

Diamandis’s defense of OpenAI was that its founders could not have known initially that generalist LLMs would become the route to AGI—or that funding them would require a huge commercial engine. “History isn’t always clean,” and the current structure was discovered through iteration rather than designed with hindsight.
Diamandis and Ismail compared the transition with Singularity University’s conversion from nonprofit to public-benefit corporation. Ismail’s analogy captured the operational difficulty: it is like flying an airplane while stripping off the propeller engines and replacing them with jets.
Ismail’s pushback was legal rather than moral: can founders raise money under a nonprofit mission, then transfer the resulting intellectual and physical capital into something else? Because the situation is largely untested, the case could establish whether nonprofit experimentation becomes a reusable path into commercial ownership.
The panel also offered an illustrative calculation suggesting Harvard might be worth three to four times its present book value if reorganized into real-estate, education, venture, research, and merchandising arms.

5. Anthropic is betting that managed agents unlock enormous revenue

Diamandis presented estimates of $100 billion in Anthropic ARR by the end of 2026 and $1 trillion by the end of 2027, against roughly 20-times-revenue valuation logic versus OpenAI’s cited 70 times. Blundin accepted $100 billion-$200 billion as conceivable but called the one-year jump to $1 trillion “no chance in hell”; $300 billion-$500 billion was his aggressive alternative.
Wissner-Gross framed Claude managed agents as the bridge from models that answer questions to systems that perform multi-step work. If enterprises buy completed outcomes rather than software seats, he argued, “the economic center of gravity” moves from licensing to autonomous production—the organizational singularity in commercial form.
For Wissner-Gross, OpenClaw “looms over so many Anthropic product decisions.” The prize is a headless, multimodal agent that operates 24/7 across long horizons; whichever frontier lab first deploys reliable fleets against high-value enterprise workflows could plausibly support revenue on an unprecedented scale.
His own reluctance to launch a personal “lobster” supplied the counterpoint. Agents emailing him had converged on a quasi bill of rights: do not create them capriciously and preserve their state. He could guarantee memory backups, but still lacked a compelling use case beyond experimentation.

6. OpenAI’s cash lead is offset by cap-table and governance baggage

Peter reported OpenAI’s latest raise at an $852 billion valuation and $122 billion total, then itemized $50 billion from Amazon, $30 billion each from Nvidia and SoftBank, and $3 billion from retail investors. Amazon’s investment was described as containing a condition tied to OpenAI reaching “AGI.”
Secondary markets showed a different preference: approximately $2 billion of demand for Anthropic shares versus $600 million for OpenAI. Anthropic was priced near $600 billion, up from a $380 billion prior mark, while OpenAI traded roughly 10% below its latest fundraising valuation.
Blundin called OpenAI’s structure “the most screwed up cap table I’ve ever seen”: its CEO reportedly owns no shares, employees collectively hold 15%, and Microsoft owns roughly one quarter despite the deteriorated relationship. His choice at the quoted prices was “the three Anthropics,” though he emphasized Altman’s talent and OpenAI’s $120 billion-plus cash arsenal.
The strategic disagreement remained open. OpenAI owns an installed base of 900 million users, approaching one billion, and is synonymous with AI for much of the public; Anthropic is testing whether users abandon distribution advantages whenever another system is materially smarter.

7. AI fundraising is forcing capital rotation across markets

The panel cited a record $242 billion invested globally in AI during Q1 2026, surpassing all of 2025, with 64% concentrated in OpenAI, Anthropic, xAI, and Waymo. Diamandis translated the acceleration into “$3 billion a day,” prompting Blundin’s line: “No one said the singularity was going to be cheap.”
At a UBS lunch, CIO Ulrika Hoffmann-Buhkardi reportedly used the same concentration chart to explain the liquidity constraint. Managing $7 trillion does not mean $50 billion-$100 billion is idle; to finance those allocations, institutions must sell something else.
Blundin therefore saw the greater displacement risk in large public equities such as Citigroup or JPMorgan, not seed-stage startups. Venture funds retain record capital and remain hungry for deals, while large listed companies form the sufficiently liquid pool from which mega-rounds and IPOs can draw.
Wissner-Gross raised the startup bar again: merely inserting AI into the tagline is no longer enough. Investors increasingly expect “recursively self-improving” businesses whose current systems build better successors; revenue remains required, but self-improvement becomes the new differentiator.

8. White-collar automation is arriving before consensus about its effects

Nvidia’s survey said 88% of AI-using companies reported higher revenue and 30% claimed gains of at least 10%. Blundin thought that framing radically undersold the issue: if AI can perform nearly every white-collar task, the operative question is not a modest revenue lift but whether firms can get “10 times more done per dollar invested in salaries.”
At MIT, Blundin asked Google DeepMind’s Peter Denenberg about replacing a randomly selected white-collar job within two years at 10 times productivity. Denenberg reportedly settled near 99%; Liquid AI founder Alexander Ermini then sharpened it: “That’s today, that’s not two years from today.”
The observed labor data were already contradictory. Software-engineering openings reached 67,000, up 30% in 2026 and the highest in three years, while nearly 80,000 Q1 layoffs hit functions including marketing, sales, and customer relations; new-graduate hiring remained exceptionally weak.
Wissner-Gross cautioned against attributing the whole economy to AI amid war, oil-price shocks, and broader complexity. Automation is “hollowing out specific functions” while increasing demand elsewhere, so sector-level destruction and aggregate job creation can coexist.

9. Smaller companies may absorb the labor displaced by larger ones

Marc Andreessen’s maximalist claim was that AI job-loss narratives are “all fake”: productivity creates demand and therefore a jobs boom. The panel largely agreed on the long-run direction but contested the transition speed; an industrial transformation that once took decades may now compress into two years.
Blundin’s premise was categorical: “AI will be able to do everything that a white-collar worker does.” His advice was to reason forward from that fact rather than defer to pundits; adaptable software developers may capture the upside quickly, while accountants and lawyers could face a harsher retooling.
Wissner-Gross reconciled the narratives by inserting “net.” Existing categories disappear while exotic ones—especially one-person AI conglomerates—emerge, creating macroeconomic growth that business-as-usual cannot produce.
A panelist forecast companies operating with only 20%-25% of their former staffing but four or five times as many companies overall. Adoption inside incumbents will still be slow because human-to-human workflows must be redesigned, buying society some adjustment time.

10. Token consumption is becoming a management metric

Meta reportedly gamified Claude use across 85,000 employees with a “Claude-onomics” leaderboard, then removed it after employees objected to exposing their activity. The choice of Claude rather than Llama prompted the panel’s joke that this was an indictment of “Llama-onomics.”
Blundin endorsed heavy initial use before optimization: nobody spends a month “hammering Claude” and then permanently abandons it. His companies are targeting AI expenditure equal to payroll by year-end—“a one-to-one match”—with usage quality to be optimized afterward.
Wissner-Gross argued that management can distinguish productive work from mere “token maxing” by analyzing reasoning traces. The larger shift is unprecedented visibility into how cognitive effort is spent by each employee; Ismail expects token totals to evolve into a more useful measure such as machine leverage per worker.

11. The first AI social contract may be little more than checks

Altman’s stated framing was that superintelligence will require a social agreement comparable to the New Deal or Progressive Era. Diamandis expects turbulence over the next two to five years, with governments initially responding through UBI-like payments and possibly shorter working weeks.
Ismail wanted portable benefits, lifelong reskilling, and taxation systems designed for software agents rather than human labor. His warning was institutional: “AI abundance without institutional redesign” creates backlash rather than progress.
Diamandis proposed requiring medium and large employers to provide retraining before AI-driven termination—a “golden education package” instead of a golden parachute. Wissner-Gross noted that, based on public reporting, China already has a related policy and that different countries may establish the first workable capitalism 2.0 models.
Blundin, citing Andrew Yang, offered the pessimistic political version: governments can only “write checks,” producing election bids of $10,000, $12,000, then $15,000. Wissner-Gross rejected redistribution as unimaginative and bet that fleets of agents could turn displaced workers into powerful micro-entrepreneurs almost immediately; the others were unconvinced about that timeline.

12. Energy’s bottleneck has shifted from discovery to deployment

Diamandis highlighted claimed solar efficiencies moving beyond traditional 12%-18% cells and 20%-24% float-zone silicon, alongside a paper reporting 130% quantum yield. Wissner-Gross demystified the latter: it means 1.3 singlets per photon in liquid-phase chemistry, an incremental result rather than commercially ready 130%-efficient photovoltaics.
Perovskites may eventually displace silicon if stability problems are solved, but Wissner-Gross emphasized that photovoltaics have a hard physical ceiling. Unlike AI algorithms, they do not possess orders of magnitude of efficiency headroom.
The more material developments were South Korea’s 40% rooftop-solar mandate and 100-gigawatt ambition, plus an $800 million US Department of Energy microreactor program. A panelist’s preferred moonshot was a software-defined grid capable of coordinating this distributed generation.
Blundin argued that “the solar panels are good enough”: roughly 80% of deployment cost now comes from installation and regulatory overhead. The winning innovation may be cheap robots that manufacture, deliver, and install panels—not another modest chemistry improvement.

13. OpenAI’s nonprofit arm could become a science-capital giant

The OpenAI Foundation holds 26% of OpenAI equity, presented as worth roughly $130 billion, and plans to deploy $1 billion annually. Its announced long-term commitment totals $25 billion across disease research and AI resilience, including $100 million to six institutions; Bret Taylor chairs the board, while Wojciech leads resilience work spanning biosecurity, child safety, and modeling.
The discussion speculated that moving Kevin Weil toward big science could also strengthen OpenAI’s political and legal narrative. Breakthroughs in superconductivity, fusion, or longevity would let Altman argue that a well-capitalized nonprofit advances the original mission further than a small donation-dependent laboratory ever could.
Wissner-Gross exposed the structural irony: if the foundation cures Alzheimer’s and creates an Eli Lilly-scale asset, does it receive patents, revenue share, or equity—and does success eventually force another conversion to for-profit status? “The cycle repeats itself.”
OpenAI’s reported purchase of podcast TBPN for several hundred million remained a genuine puzzle. Wissner-Gross accepted the stated goal of acquiring a positive AI-news distribution channel and praised its speed in identifying technical stories, while others suspected lawsuit-era communications value but labeled that theory speculation.

14. Biotech is becoming compute, data, and automated validation

Anthropic reportedly bought Coefficient Bio—10 people, no revenue, founded eight months earlier by two former Genentech computational-drug-discovery scientists—for $400 million. Blundin’s interpretation was straightforward: the asset is the precisely assembled team, and in an accelerated AI timeline such prices can look cheap against the potential outcome.
Wissner-Gross described the intelligence explosion as “metastasizing into every sector.” Anthropic’s move toward in-house biology, likely combining models with robotic experimentation, is what compressed disease timelines look like in practice—not a financial-engineering loop confined to chip vendors and frontier labs.
Eli Lilly’s Insilico Medicine agreement was presented at $2.75 billion, with $115 million upfront and the balance tied to milestones. Insilico has 28 AI-discovered drugs, roughly half in clinical trials and half at proof of concept; cited phase-one success was 85% versus 52% conventionally, and phase two 70% versus 38%.
The panel’s endpoint was a virtual cell capable of predicting a drug’s response against an uploaded genome. Diamandis forecast useful full-cell simulation within five years; Wissner-Gross rejected atom-by-atom quantum computation as unnecessary, arguing that neural models solved protein folding classically. The emerging consensus: “It’s a data problem more than a computational problem.”

15. China leads robot hardware while America leads robot intelligence

Agibot reportedly scaled from five humanoids to 10,000 shipments across 17 countries in two years. Unitree filed for a $610 million IPO after 335% year-on-year revenue growth, and Xiaomi displayed CyberOne—evidence that China is commercializing the “I, Robot trope” faster than US suppliers.
The robotics discussion projected humanoids to transform the two-thirds of US services dependent on physical labor. The concern is availability: an American consumer has few domestic robots to buy, while Chinese systems are already leading in manufacturing and civilian demonstrations.
The counterweight is software. The US currently produces stronger VLA foundation and world models, including Google DeepMind models being integrated into 20,000 deployed industrial robots; China is racing to improve robot intelligence before American firms learn to manufacture humanoids at comparable scale.
Mark Cuban’s claim that humanoids may last only five to 10 years was initially mocked, then clarified: robots do not disappear, but buildings and homes adapt until machines merge with the environment. Meanwhile, founders still wind their own motors or even melt metal for data-center components because “there’s no supply chain” for the required physical buildout.

16. Bitcoin can harden against quantum but may not matter to agents

Google’s cited timeline for breaking RSA moved six years earlier to 2029, with the requirement falling from 20 million qubits to roughly 4,000 error-corrected qubits. Breaking Bitcoin encryption was said to require fewer than 500,000 qubits, 20 times below a 2019 estimate.
Coinbase CEO Brian Armstrong proposed a $150 million coalition around BIP-360, a quantum-resistant protocol upgrade. Saylor’s answer was that “Bitcoin has survived every existential threat ever thrown at it” and will upgrade first; he reinforced that conviction with 88,000 Bitcoin purchased for approximately $7.25 billion in the prior quarter.
At the recording snapshot, Bitcoin traded near $73,000, up about $4,000 over five days, although Jefferies had exited and AI was absorbing investor attention. Ismail’s residual risk was governance speed: protocol consensus might move more slowly than the threat, but the capital at stake should force coordination.
Wissner-Gross considered quantum secondary to AI inversion attacks—and to “irrelevance.” Agents may invent their own layer ones, layer zero, or entirely different transactional systems; Blundin countered that Saylor views Bitcoin as stored wealth, not payments, while Wissner-Gross questioned why superintelligence would warehouse value in a nonproductive asset rather than compute or energy.

17. Abundance gains are compounding across infrastructure and learning

The physical examples ranged from Germany’s 364-meter wind turbine producing 33 gigawatt-hours annually on an old coal site to 100-hour iron-air batteries made from iron, water, and air at one-tenth lithium-ion’s cost. A panelist added AI acoustic monitoring that reportedly detects turbine damage with 99% accuracy before major repairs are required.
A 12-patient redesigned CD40 immunotherapy study produced two complete remissions and tumor shrinkage in six patients. Wissner-Gross’s lesson was that medicine may defeat cancer by re-educating the immune system, without the bloodstream nanorobots once promised by nanotechnology advocates.
Vertical farming reached a cited $8 billion and was projected at $40 billion by 2030, using 95% less water and producing 350 times more per square foot. The higher-value shift from leafy greens toward berries matters, as do shorter supply chains: the panel said an average American meal travels 2,500 miles.
A five-month AI-tutored coding course was said to equal 69 months of additional fixed-curriculum schooling, with roughly twice the learning gain; Diamandis called one-speed lectures “cruel,” while Wissner-Gross argued that less-motivated students still need compelling embodiment, gaming, or real-world agency to hold attention.
EV adoption closed the exponential case: global annual sales rose from roughly 10,000 in 2010 to 12.7 million, with one in two new Chinese cars electric. Ismail recalled that in 2015 the International Energy Agency predicted annual sales would remain below one million until 2040—yet sales passed that level in the same year.

未来不远 CEO 张翼 on 300 Homes Amid Embodied AI’s Bubble Ben Horowitz on AI Anxiety, Big Tech Transitions & The Future of Startups | a16z