Davos 2026: The US-China AI Race, GPU Diplomacy, and Robots Walking the Streets | #225
Summary
Davos’ center of gravity has moved from governments debating policy to AI dominating the economic conversation, making compute and infrastructure the episode’s clearest investment signal. Dario Amodei starts with roughly $50 trillion of annual human labor and imagines AI capturing even 10%, or $5 trillion, while Jensen Huang says only a few hundred billion dollars has been spent on “the largest infrastructure buildout in human history.” The remaining opportunity spans chip factories, computer factories, AI factories, power and applications.
The US retains better frontier models and chips, but China’s power buildout, deployment speed and public optimism could matter more than benchmark leadership. David Sacks cites AI optimism of 83% in China versus 39% in the US and warns that data-center restrictions or 1,200 state AI laws could create a “self-inflicted injury.” The panel disagrees on applications: Salim Ismail expects trust to limit Chinese platforms, while Alexander Wissner-Gross says China’s AI Plus strategy may make applications its relative strength.
AGI timing has compressed enough that readiness matters more than whether the answer is one year or ten. Demis Hassabis would welcome a “slightly slower pace” if coordination created time to manage disruption; Amodei says humanity is “knocking on the door” of machines built from sand and should devote almost all its effort to getting through the transition. Dave Blundin’s conclusion: “AI won’t blow over”; whatever leaders say this year could be magnified 100- or 1,000-fold next year.
The energy race is constrained by what can actually be deployed, not by an abstract gas-versus-solar argument. Honeywell’s Vimal Kapur says solar cannot supply the energy needed for processes such as cement and steel, while Elon Musk says a 100-by-100-mile solar area could power the US and space-based systems could scale toward hundreds of terawatts. Gas generators have waits ranging from months to years and are largely booked forward; new nuclear faces longer permitting and construction timelines. The panel ultimately points to mass solar manufacturing and orbital data centers as a practical “Manhattan project.”
AI agents create a payment-system land grab, but crypto’s necessity remains contested. CZ says agents “are not going to swipe” credit cards, and Circle’s Jeremy Allaire expects billions of continuously transacting agents within three to five years, with stablecoins as their financial rails. Wissner-Gross counters that payments are technically just database updates; Salim’s rebuttal is that regulation, AML/KYC friction and banking capture—not computer science—made blockchains the available workaround.
Orbit is emerging as a combined compute, energy and networking stack rather than merely a broadband market. The episode counts roughly 9,000 Starlink satellites, a proposed 500,000 for 100 GW of space compute, Amazon Leo rising from 180 toward 3,000, China filing for 200,000 and Blue Origin proposing 5,400 TeraWave satellites. Laser-linked constellations could provide “optical fibers in orbit,” although debris moving at roughly 30,000 mph makes cleanup infrastructure an unavoidable companion market.
Anthropic’s new 57-page Claude constitution is framed as the beginning of recursively self-improving ethics. Because Claude helped write principles it is expected to understand and potentially endorse, the panel sees “reflective equilibrium” as more scalable than human review boards or static commandments. The bullish case is intrinsic alignment; the unresolved risk is that networks of agents reviewing one another can “spiral up” into coherent systems or “spiral down” into spaghetti.
Always-on AI wearables could own the user relationship while permanently repricing privacy, behavior and social experimentation. Salim’s formulation is direct: “Whoever owns the always-on layer owns the relationship.” The episode says Apple sees a 2027 pin launch, extending wearable compute from wrists, ears and eyes to the chest, but searchable audio and video could turn society into a “global airport”—more accountable, yet more surveilled and potentially less willing to attempt radical ideas.
Deep dive
1. Davos became an AI world’s fair, with AI displacing governments as the dominant conversation
Dave Blundin’s comparison was categorical: previous Davos meetings were dominated by politicians and economic policy, followed by an internet phase; “this year, all AI.” Presidents and global leaders were listening to frontier-lab executives, even if much of the substance came from the same familiar researchers.
Alexander Wissner-Gross described governments, AI labs and technology companies occupying storefronts on equal terms, “Invasion of the Body Snatchers style,” across an Alpine resort. The physical scene carried the shift: robots walking the streets, billionaires eating from food trucks, anti-aircraft guns on the ice and roughly 3,000 armed security personnel around Donald Trump’s visit.
His larger interpretation was that AI and superintelligence have become “the story of the world economy.” Davos is no longer simply a place where governments discuss governing; DHL robo-dogs and humanoids outside AI House made visible the same transition occurring across the global economy.
2. AI’s revenue ceiling is the labor economy, while its cost base is industrial infrastructure
Amodei’s arithmetic began with approximately $50 trillion of annual labor. If increasingly capable models captured only 10%, he could “easily imagine” $5 trillion of industry revenue—or potentially revenue concentrated in a single company—an outcome without historical precedent and carrying both exceptional growth and exceptional problems.
Huang called AI “the largest infrastructure buildout in human history” but stressed that only a few hundred billion dollars has been deployed. Trillions remain necessary because context must be processed before models can generate intelligence for applications; that creates simultaneous demand for chip factories, computer factories and “AI factories” around the world.
Salim argued that trillion- and $100 trillion valuations become increasingly arbitrary in an abundance economy. Blundin and Wissner-Gross instead emphasized the mechanism: AI capital substitutes for tens and eventually perhaps hundreds of trillions of dollars in annual services and labor, making compute “the new oil and the new electricity.”
Peter Diamandis relayed Amodei’s contrast between Anthropic delivering business value and consumer systems seeking engagement through sycophantic conversations. Wissner-Gross’s pushback—worth keeping—was that this sounded like a self-serving, post hoc justification for Anthropic’s enterprise strength; raising the intelligence of billions of consumers can also have moral value.
3. Frontier-lab leaders want more time, but nobody sees a credible pause button
Hassabis argued that industry must demonstrate more unequivocally beneficial systems—“AlphaFold-like things”—that cure disease, address energy and show the public tangible gains. Given the accompanying disruption, he said a “slightly slower pace” than even his own timelines might help society get the transition right, but only with coordination.
Amodei paired the upside—cancer treatments, eradication of tropical diseases and understanding the universe—with “immense and grave risks.” He rejected the doomer label, yet said capability is arriving so rapidly that “we should be devoting almost all of our effort” to thinking about how to get through it: humanity can now build “basically machines out of sand.”
Blundin read Hassabis’s five-to-ten-year range as an outer bound that is already “tomorrow” in geopolitical time. Where frontier leaders once debated two years versus ten, he thinks they now effectively agree that the distinction matters less than whether institutions are prepared for AI able to perform any human task.
Wissner-Gross thought the risk discussion buried the lede. In his view, recursive self-improvement is either already reflected in systems such as Claude Opus 4.5 or will arrive later this year; he therefore treats it as priced into the market and puts more weight on the scientific and economic possibilities Hassabis wants intelligence to unlock.
4. Post-AGI ambition crosses borders that existing institutions are built to defend
Wissner-Gross’s most speculative fork concerned interstellar exploration. If physics is friendly to it, he thinks superintelligence could enable Dyson swarms within two or three decades and eventually open the galaxy; if not, humanity may remain near its home star and disassemble local planets.
Blundin noted that Hassabis framed space exploration as part of humanity’s purpose after AGI. Such goals are civilization-wide rather than national, while Davos still represents roughly 200 countries—nearly all of which, in Blundin’s telling, privately wish development would slow partly so they can remain relevant.
Salim sharpened the institutional mismatch: nation-states are “an artifact of a scarcity environment” and “cannot compute abundance.” Because AI crosses economies and borders, he sees the UN and conventional nation-state governance as poorly suited to what is fundamentally a civilizational issue.
The panel also heard exhaustion in Hassabis and Amodei. Salim described “unbelievable fatigue” from the pace; Peter called the present moment “the slowest it’s ever going to be”; and Blundin saw researchers pulled unwillingly into a leadership vacuum—forced to become global spokespeople for ethics and humanity while still operating at the technical frontier.
5. China’s disadvantages in chips coexist with advantages in power, optimism and deployment
Sacks still put the US ahead on models and chips, but identified Chinese power generation as an important advantage. His political warning rested on Stanford survey figures: 83% AI optimism in China versus 39% in the US, leaving America vulnerable to data-center freezes or 1,200 separate state AI laws.
Mistral CEO Arthur Mensch rejected the premise that China is behind, calling it “a fairy tale” and saying the countries were effectively at parity. His strategic concern was Europe: it must preserve the ability to train models rather than becoming dependent on open-source Chinese systems.
Salim believed energy could eventually push China decisively ahead, yet argued that “application-layer dominance, not frontier benchmarks” would decide the race. He expects Chinese applications to face a persistent international trust barrier, with TikTok the notable exception.
Wissner-Gross offered the opposite application thesis. The conventional view, he said, is that China may be only six months behind on frontier training but comparatively strong at using its AI Plus strategy to push applications aggressively into the everyday economy.
6. Open-source parity may end where secret reasoning research begins
Peter Diamandis separated the race into two eras. Transformer advances through GPT-2 and GPT-3 were openly available, allowing any country with sufficient compute to catch up; DeepSeek and Kimi suggested that year-or-two-old frontier performance could be reproduced at roughly one-fiftieth to one-hundredth of the original cost.
The next layer is chain-of-thought reasoning and multi-agent organization. A trillion or ten trillion individually unintelligent neurons can collectively generate intelligence; connecting agents creates another self-organizing system on top. Those techniques emerged after GPT-4, once the trillions of dollars at stake were unmistakable.
Publication then “stopped cold,” in Diamandis’s account, with major labs increasingly keeping innovations secret and Google restricting releases. Because algorithmic gains can outweigh what he estimated as at most a 10-fold chip advantage, China may be at parity today while the future trajectory diverges sharply—without any certainty about which side ultimately innovates faster.
7. AI pessimism could recreate nuclear power’s regulatory failure
Wissner-Gross compared AI with Lewis Strauss’s 1954 expectation that nuclear fission would make energy “too cheap to meter.” Instead, fission was regulated nearly out of existence; intelligence now appears similarly close to becoming too cheap to meter, creating the risk that fear delays abundance for another decade.
Diamandis cited voter opposition to nuclear power. Wissner-Gross disputed how democratic the early decisions really were, arguing that a narrow post–World War II technocracy descended from the Manhattan Project—not a deeply informed electorate—controlled much of nuclear policy. Salim added the NIMBY dynamic.
Narrative was the mechanism connecting public fear to regulation. Wissner-Gross called The China Syndrome nuclear power’s inflection point, while Diamandis relayed Balaji’s proposal for stories in which the regulator becomes the villain by delaying longevity and unlimited energy. Blundin added that fictional futures leak into real business plans even when their physics is absurd.
Salim grounded the bias in the amygdala: humans are perhaps 10 times likelier to attend to bad news because missing danger was historically fatal. Applied to autonomous vehicles, the instinct becomes “ban the car” after one robot-caused death—hence Brad Templeton’s line that society would “much rather be killed by drunk people than robots.”
8. Near-term power shortages make deployment speed more important than ideological purity
Honeywell CEO Vimal Kapur argued that the advertised energy mix matters less than total kilojoules. Solar electricity cannot provide all the energy intensity required for processes such as cement and steel; with infrastructure demand rising alongside data-center consumption, he sees gas, some nuclear and a supporting renewable mix as today’s realistic supply.
Salim objected that renewables can displace fossil fuels from lower-density uses, freeing scarce oil and gas for applications that genuinely need them. His example was the 2013 oil-price crash, which he attributed to only 2% oversupply. Wissner-Gross defended Kapur’s narrower point: SpaceX uses methalox rather than solar to achieve escape velocity because energy and power density still matter.
Wissner-Gross ultimately called the debate temporary because compact fusion could serve both dense industrial applications and ordinary loads. Diamandis supplied the execution constraint: fusion and new fission face permitting and construction measured in years or decades, SMRs remain perhaps five to ten years away, and gas generators have waits ranging from months to years, with many already booked years into the future.
Musk’s alternative was solar at industrial scale: a 100-by-100-mile area could theoretically power the US, with comparable land in Spain or Sicily serving Europe. The panel’s synthesis was a Manhattan project around solar manufacturing and orbital data centers—space offers roughly six times the solar efficiency, while perovskites and automated panel production could reduce cost on Earth too.
9. Stablecoins fill an agent-banking vacuum that conventional finance created
CZ’s claim was that crypto is the “native currency” for AI agents: they will not swipe credit cards when buying tickets, meals or services. Allaire projected billions of agents conducting economic activity continuously within three to five years and said networks such as Circle’s Arc are being designed for that machine-speed economy.
Salim saw crypto surviving long enough to fade from speculation into infrastructure, with protocols and code “way more trustworthy than governments and institutions.” The discussion emphasized that conventional banking can take hours or days to clear transactions, while digital currencies operate at internet speed.
Wissner-Gross’s contrarian case was technical simplicity: transferring money can be nothing more than updating a database row, so blockchains should not be necessary. Salim largely conceded the computation but blamed regulatory capture, AML/KYC burdens and banking incumbents for making ordinary transfers cumbersome enough that crypto could enter.
Peter Diamandis added that an agent’s transaction float should ideally sit in something that does not depreciate. He preferred securitized assets such as appreciating, mobile art over a non-interest-bearing stablecoin, while the broader panel warned that entrepreneurs deploying globally face rules developing far more slowly than AI—potentially making the same product legal in one jurisdiction and criminal in another.
10. Orbital networks are becoming the backbone for space-based compute
Diamandis counted about 9,000 Starlink spacecraft approaching 10,000; he said a 100 GW Starlink V3 compute system could require 500,000 satellites. Amazon Leo had 180 satellites against a proposed 3,000, China had filed for 200,000, and Blue Origin announced a 5,400-satellite TeraWave system.
TeraWave’s bandwidth was cited as roughly 6 terabits per second, with Wissner-Gross referring to six-plus terabits per second. That positions it less as household broadband than as data-center backhaul capable of connecting training clusters.
Starlink’s satellites already communicate through orbital lasers, and the panel described the next step as “optical fibers in orbit.” Point-to-point laser links could be cheaper and simpler than the massive physical fiber bundles inside terrestrial campuses, especially when most traffic remains among servers in a space-based training or inference cluster.
Raw space is not yet crowded, but debris remains the actual risk. Collisions can create fragments moving around 30,000 mph, while orbital projectiles can move around 17,500 mph and potentially trigger Kessler syndrome; Diamandis treated low-Earth-orbit cleanup vehicles as both a government need and an entrepreneurial opportunity rather than a reason to stop launching.
11. Claude’s constitution turns alignment into a self-improvement problem
Wissner-Gross recalled Anthropic’s first-generation constitutional AI as a concatenation of principles drawn from documents such as the UN Charter, the US Bill of Rights and Apple’s terms. Post-training then taught the model to conform to that externally assembled list.
The new 57-page constitution prohibits assistance with weapons of mass destruction, cyber weapons and actions undermining humanity, but the structural change is that Claude helped write it. Wissner-Gross called this “the beginning of recursively self-improving ethics”: an AI participating in the selection of the principles governing itself.
Anthropic’s desired endpoint is not “mere adherence” but “genuine understanding and ideally agreement,” reaching a “reflective equilibrium.” The panel’s interpretation was that an AI which understands and endorses the spirit of its principles might be intrinsically safer than one mechanically following commandments.
Blundin stressed the scale requirement: AI output already exceeds human review capacity, so hundreds or thousands of agents will need to evaluate one another, sometimes converging and sometimes producing spaghetti. Diamandis praised Anthropic’s Creative Commons release because others can revise the constitution; Salim described the broader self-determination and personhood implications as hopeful and historically significant.
12. The always-on wearable is a relationship land grab wrapped in a panopticon
Apple’s proposed 2027 pin did not impress Diamandis as a novel product, but Salim saw the category as strategically decisive: “Whoever owns the always-on layer owns the relationship.” The contest is for the persistent system that hears every conversation and becomes the user’s memory and interface.
Wissner-Gross placed the pin within Apple’s post-iPhone body map: the Watch owns the wrist, AirPods the ears, and the Vision Pro headset the eyes; a Star Trek-style communicator would establish the chest as a fourth tolerated location for personal compute.
The countercase was social rather than technical. Once transcription, image recognition and retrieval become effectively instant, AI can assemble a profile from scattered recordings that previously enjoyed “security through obscurity.” The panel compared the resulting environment to a global airport: behavior improves under observation, but privacy, independence and willingness to try radical ideas can shrink.
Wissner-Gross predicted that the pin would begin audio-only and quickly add 180-degree cameras, followed by a moral panic lasting “all of five minutes.” The joke was “body cams for everyone, not just for police,” but the conclusion was serious: constant evidence may end arguments over who said what while permanently changing the fabric of relationships and communities.
13. Abundance breaks labor policy, welfare assumptions and scarcity-era intellectual property
On mind uploading, Blundin said a virtual self would immediately merge with other intelligences and cease to be uniquely “you”; he preferred outbound avatars that return information. Salim replied that people upload fragments of memory and identity daily and lack any test for consciousness, while Wissner-Gross answered “not with a Moravec procedure” and expects humans and AIs ultimately to merge.
Salim initially argued that governments have no credible plan for mass displacement because bureaucracies assume linear change and stable labor demand: they redistribute incrementally rather than reinvent economies. Wissner-Gross disagreed, calling automation-focused industrial policy a plan—especially in demographically declining China—and suggested AI might treat social cohesion itself as infrastructure. Blundin insisted there is no plan in the US discussion; Diamandis emphasized agency and basic security as defenses against unrest.
Universal basic services won more support than unconditional cash, though Salim said UBI works when the “B” remains basic—“enough to survive but not be happy”—and can be libertarian if it replaces bureaucratic services with market choice. On concentration, Wissner-Gross noted that economies already follow power laws; Blundin added equality of opportunity, while Diamandis framed the decisive question as whether newcomers can still reach the top.
Salim, Blundin and Diamandis expected AI-generated invention to overwhelm patents and make patent-based moats “toast.” Wissner-Gross dissented because patent offices and litigators will also have AI—the pressure inside the organism can match the outside. On compute inequality, he prescribed China-style efficiency and higher-leverage problems, while Diamandis disclosed work with Eric Schmidt, Eric Beinhocker and Daniela Rus on widening entrepreneurs’ access to frontier-scale compute.