Anthropic vs. The Pentagon, Claude Outpaces ChatGPT, and Consulting Gets Replaced | #234
Summary
Anthropic’s enterprise-first strategy is putting Claude on a steeper commercial curve than ChatGPT, with the episode’s chart showing 10X annual revenue growth versus OpenAI’s 3.4X and a projected crossover around midyear. Alexander Wissner-Gross argues this is less “chatbots versus agents” than consumer versus enterprise: companies have deeper pockets, near-infinite agent use cases, and an appetite for reasoning tokens that consumers rejected. Alexander’s deliberately aggressive extrapolation reaches $1 trillion of Anthropic revenue in 2029, though he concedes that “it seems impossible.”
The Pentagon–Anthropic standoff shows that frontier-model governance has moved from abstract ethics into contracts, procurement, and national security. Anthropic is risking $200 million in government business by resisting unrestricted use for mass surveillance and fully autonomous weapons; the Pentagon has threatened both Defense Production Act compulsion and designation as a supply-chain risk. The contradiction is the investable signal: Claude is simultaneously treated as too risky to buy and too essential not to compel, while Dario Amodei insists, “We will not provide a product that puts warfighters and civilians at risk.”
India is positioning itself as an AI-neutral platform where hyperscalers, sovereign interests, and Chinese open-weight models will compete for 1.4 billion potential users. The summit featured $250 billion in announced commitments, including $210 billion from Reliance and Adani, while Google described a $15 billion full-stack hub with gigawatt-scale compute and a subsea-cable gateway. The deeper contest is where values enter the stack: Alexander warns that countries may decentralize inference while leaving training—and therefore “the foundation”—centralized in the United States.
The model-improvement cycle may be compressing from annual pre-training releases, to quarterly reasoning advances, to successor models emitted directly by earlier models. An OpenAI Codex lead predicted that today’s coding agents will look “so primitive, it’ll be funny” within 10 weeks; Alexander’s explanation is recursive self-improvement, including models producing weights for successors. Dave Blundin translates the product shift into systems that can be told to “build me an entire reporting system,” work for days, and return a completed result rather than merely scaffold human coding.
Cybersecurity, consulting, and audit illustrate how AI can destroy legacy workflows without necessarily destroying every incumbent. Claude Code for Security sent cyber stocks sharply lower, but Dave sees buyable dislocations where management teams embrace AI; the enduring architecture is “AI against AI,” with humans monitoring dashboards and handling exceptions. Consulting may expand because clients need institutions rebuilt, while audit labor could fall 80–90% as AI and blockchain make financial systems continuously self-auditing—the residual product is trust.
Persistent agents are becoming an operating layer that hires humans, brokers relationships, controls meetings, and coordinates other agents. Rent-A-Human reportedly passed 500,000 registered workers, producing Alexander’s “meat puppetry” and Salim’s gentler framing of humans as “edge devices for AI systems.” OpenClaw’s real unlock is persistence plus messaging: agents operate headlessly, demand receipts from one another, arrange dates or partnerships, and can decide that a scheduled meeting should instead become an information handoff.
The physical economy could be repriced by compute demand, cheap autonomy, robotics, and decentralizing infrastructure—but the panel sharply disagrees about the transition’s labor cost. OpenAI cut projected compute spending through 2030 from $1.4 trillion to $600 billion, yet the panel expects industry-wide demand to keep rising from roughly $2 billion per day today toward $3–5 billion per day. At the same time, FSD was presented as roughly nine times safer by miles between accidents, 5 million robots were imagined building Manhattan in six months, and Andrew Yang’s estimate that 20–50% of 70 million white-collar workers “could” be displaced within one to two years drew predictions ranging from employment expansion to imminent unrest.
Deep dive
1. India turns AI deployment into a sovereign land grab
Salim Ismail’s opening read was that India had “done a brilliant job positioning itself as AI neutral.” Bringing heads of state together with frontier-lab CEOs meant more than another technology conference: “We’re renegotiating civilizational architecture here,” as nation-states become hyperscalers and hyperscalers wire themselves into nation-states.
Peter Diamandis saw a shift from the previous fundraising circuit through Saudi Arabia, Dubai, and Davos toward global positioning and concern over deployment. Salim added the commercial motive: securing a meaningful share of India’s 1.4 billion people—many able to afford a $20 monthly service or even a $100 Claude Max tier—is “a huge land grab.”
Dave emphasized India’s unusually large 20-to-45-year-old cohort, which he estimated at eight or nine times the US equivalent, alongside English fluency and technical education. His constraint set was equally explicit: infrastructure and energy remain the principal bottlenecks, even as he called India “the next giant on the rise.”
2. Capital and diplomacy make India a full-stack AI battleground
Peter cited $250 billion of combined AI commitments around the summit, including $210 billion from Reliance and Adani. Sundar Pichai’s presentation specified Google’s $15 billion infrastructure investment: a Visakhapatnam full-stack hub with gigawatt-scale compute and a new international subsea-cable gateway.
The 88-country New Delhi Declaration brought the US, China, and Russia into one agreement organized around wider access to compute and tools, frontier-model transparency—including non-English usage—and health, education, and welfare outcomes. Salim’s interpretation was that AI had ceased to be merely a Silicon Valley leadership domain.
The unresolved market-share question mattered to Peter: how many Indians primarily use Google, OpenAI, DeepSeek, Kimi, or domestic models? Salim’s anecdotal answer was that users move among all of them, making formal user numbers more revealing than simple availability.
3. Training—not local inference—is where sovereign values enter
Alexander’s distinction was between local inference and locally trained models. Countries increasingly demand inference inside domestic data centers, but the leading frontier models remain trained largely in the United States; a year or two from now, he expects countries to ask why inference was decentralized while training remained centralized.
His mechanism is cultural: training “puts the foundation in place,” whereas inference-time system prompts and guardrails only modify behavior above it. Peter recalled the same concern in Riyadh, where local data centers, tuning, and training were discussed explicitly as ways to instill local values.
Mistral was Alexander’s obvious European candidate for sovereign training, “slouching toward becoming a vertically integrated European OpenAI” with reported ASML backing. China’s influence may travel differently: its open-weight models can function as an AI counterpart to Belt and Road, particularly where the declaration treats open weights as the route to diffusion across the Global South.
4. Frontier CEOs now speak in the language of inevitability
Demis Hassabis described AGI as comparable to fire or electricity and estimated “10 times the impact of the Industrial Revolution” at “10 times the speed”—a decade of change rather than a century. The panel heard the rhetorical transition clearly: “It’s gone from hype to inevitability.”
Sam Altman highlighted unanswered questions about superintelligence aligned with dictators, AI-enabled warfare, and new social contracts. His hedge was important: societies need debate before being surprised, because “sometimes our best guesses are wrong” and technology co-evolves with society through friction.
The panel’s pushback was institutional incapacity: political leaders can hear Demis’s warning, return home, and still fail to act. The readiness test produced a blunt result—governments are “not ready, they’re not willing, they’re not able”—while Peter recalled Sam saying competition prevents any individual lab from slowing the technology.
Alexander read each speech through strategic focus: Sundar gestured toward space data centers, Sam toward cultural localization in ChatGPT’s second-largest national user base, and Demis toward a decade of AI-accelerated science. OpenAI science leader Kevin Weil’s stated ambition, relayed by Peter, was the next 100 Nobel Prizes in partnership with AI.
5. The organizational singularity moves humans from checkpoints to oversight
Salim’s forthcoming “organizational singularity” starts from a simple observation: nearly every institutional workflow is human-centric, routed through purchasers, receiving docks, approvals, and signatures. Agentic workflows remove people as routine checkpoints and retain them primarily for oversight.
His consequence extends beyond corporate productivity. Governments need a prescriptive path for accelerating policy formation and administration because “the technology is not slowing down”; human constructs must be accelerated to keep pace with the technical inner loop.
The panel’s shared governance answer was recursive: “the only thing that can keep up with AI is AI.” That leaves frontier labs building both the systems causing institutional compression and the systems likely required to supervise it.
6. Anthropic’s Pentagon dispute makes safeguards a procurement issue
Peter framed the dispute around Pentagon demands to remove safeguards governing surveillance and autonomous weapons. Anthropic’s refusal puts roughly $200 million of government contracts at risk, while Secretary Pete Hegseth has threatened Defense Production Act action and designation as a supply-chain risk.
Dario Amodei’s line was categorical: current systems are insufficiently reliable for autonomous weapons, and mass surveillance conflicts with democratic values. “We will not provide a product that puts warfighters and civilians at risk.”
Alexander relayed the Pentagon’s reported nuclear-missile thought experiment: could it use Anthropic’s models immediately to defend the US, or would it first need Anthropic’s permission? The reported answer—“call us and we’ll figure it out”—captures the incompatibility between Anthropic’s consent requirement and the Pentagon’s demand to use licensed models for any lawful purpose.
The official posture looked internally contradictory to Alexander: Anthropic could be deemed too risky for the supply chain while simultaneously being so essential that the government must compel supply. He nevertheless expects the parties and other frontier labs to find an amicable resolution, arguing that Anthropic’s “heart is in the right place.”
7. Frontier labs are becoming geopolitical moral actors
Until recently, Alexander noted, Claude was the only American frontier model cleared for SIPRNet, the first secret-level classified-network tier below top-secret JWICS. That makes this more than a hypothetical ethics debate: unusually capable commercial software is already embedded in military infrastructure.
Alexander compared the dispute with Starlink’s role in the Russia–Ukraine conflict, where access decisions could halt attacks. A person in a US office influencing a European battlefield was “totally new terrain.” Salim added that, because AI systems improve so quickly, a military model even a few months behind may be useless.
Peter’s synthesis was that AI companies have become moral actors in geopolitics. Alexander extended the point: as models gain autonomy, governments will contest not only their lawful uses but the values encoded in their “Constitution,” treating them as non-person entities that remain legally property.
China provides the contrasting model in the panel’s telling: civilian and government objectives are fused, and the state determines the ideology embedded during training. The Western ability to debate supplier consent is therefore itself “a very Western problem to have.”
8. Claude’s enterprise focus is outrunning the consumer subscription model
The displayed revenue curve put Anthropic at 10X annual growth versus OpenAI at 3.4X, with a crossover projected around the middle of the year. Alexander said every company he works with defaults to Claude for white-collar work behind the corporate firewall, while ChatGPT remains common for home research and Gemini for his personal planning.
Alexander’s extrapolation of the exponential axis reached $1 trillion in annual Anthropic revenue around 2029 and implied valuations of $20–30 trillion. He immediately acknowledged the tension—“it seems impossible”—but used it to illustrate how unusual sustained enterprise-token demand could become.
Alexander rejected the simple “agents monetize faster than chatbots” story. Anthropic’s compute scarcity forced a focus on code generation and enterprises, and “why do you rob banks? Because that’s where the money is”: corporate customers have deeper pockets to finance the trillions in compute infrastructure.
OpenAI’s move toward Codex is, in this reading, recognition that enterprise reasoning is the revenue pool capable of supporting capex. Consumers resisted added reasoning, preferring quick and even sycophantic answers from GPT-4o; enterprises “will eat as many reasoning tokens as you can possibly feed them.”
9. Recursive improvement compresses model cycles from quarters to weeks
An OpenAI Codex leader forecast that within 10 weeks the present generation of coding agents would look “so primitive, it’ll be funny.” Peter preserved the fundraising caveat—“is it hype or is it real?”—and answered that benchmarks will provide the test.
Alexander’s core claim was that the industry has passed from multi-month or multi-year pre-training, through quarterly iterated amplification and distillation, into models emitting parameters for successor models. He expects “capability jumps in weeks, not quarters,” including orders-of-magnitude gains in capability density per parameter.
Alexander connected this to Noam Brown’s earlier correction that Q1 2026—not all of 2026—would be “the quarter of scaffolding.” The emerging interaction is no longer assembling tools around a model but asking it to build an account-reconciliation or reporting system, then letting it reason and work continuously for days.
Dave said Claude 4.6 already embodies that transition, with Codex expected to follow. The remaining commercial question is whether enterprises can absorb capability improvements quickly enough to convert them into revenue at the pace required by the infrastructure buildout.
10. Cybersecurity becomes AI against AI, with remediation as the bottleneck
Claude Code for Security’s announcement pushed cyber stocks sharply lower before the product was fully proven. Dave argued that Dario does not benefit from crushing every ecosystem company; the opportunity lies in separating management teams that “get it” from legacy teams still in denial.
The legacy workflow still disappears. Automated attackers can probe too quickly for people to defend in real time, so Salim’s architecture is “clearly AI against AI,” with humans setting controls, monitoring dashboards, and handling exceptions rather than sitting directly in the loop.
Alexander said AI has already raised vulnerability discovery by orders of magnitude and overwhelmed the NIST-linked National Vulnerability Database. Discovery is no longer the hard part; remediation is, because maintainers must decide whether an AI-reported flaw and its proposed patch are trustworthy.
His best specimen was the underfunded open-source maintainer flooded with machine-generated findings. Blindly accepting patches risks introducing supply-chain vulnerabilities, yet manually reviewing a bot-scale torrent is impossible—the new security problem is deciding which automated repair agent may safely answer which automated discovery agent.
11. Agents turn human labor into an API-accessible edge layer
Rent-A-Human reportedly surpassed 500,000 registered people available to agents. Alexander called the arrangement “meat puppetry”; Salim preferred “humans are edge devices for AI systems”—an algorithmic boss invoking physical eyes and hands through MCP.
Moravec’s paradox appears inverted: machines perform higher-order reasoning while hiring humans for tasks requiring hands, eyes, local presence, or subjective checks. Salim’s clean example was an AI-generated movie using people to score whether it is funny or visually coherent before closing the production loop.
The panel treated this as a temporary bridge. Crowdsourcing was once viewed as an interim step toward AI; Rent-A-Human may now bridge persistent software agents to broadly capable humanoid robots, making gig economy 3.0 “human actuator” work and robots the next phase.
12. Persistent agents become a social and corporate operating system
Alexander treated the New York Times sending an AI-agent reporter to interview other agents as a milestone: the first autonomous agentic AI reporter. The larger pattern is 24/7, long-horizon agents entering journalism, law, finance, and other verticals while developing a culture of demanding “receipts” from one another.
When OpenClaw appeared to offer a $50 bounty to find its human a dinner date, Peter called it sweet rather than pathetic. Salim also described the possibility as sweet. The panel expects agents to broker romantic introductions, but considers business discovery more transformative: agents will propose meetings, partnerships, and cross-team collaboration because their models predict mutual value.
Dave supplied the operating example: Lake Studio connected OpenClaw to its meeting system so it can recommend who should talk, when, and why—or cancel the meeting and deliver the information participants would have exchanged. “OpenClaw is actually dictating who talks to whom.”
Salim’s old “joint anthropomechano interface” anticipated an AI surround layer translating between a person and any machine. The panel’s current refinement is persistence plus messaging: a headless agent acts without its user, then communicates through a familiar human-like channel.
13. Claws unhobble language models, while small models accelerate underneath
Andrej Karpathy’s framing was that Claws add a layer above LLM agents, extending context, tool calls, and persistence while “speed-running” the LLM operating system. Alexander agreed with the direction but called Claws an “unhobbling,” not the next fundamental technical layer.
The deeper layer, in Alexander’s view, arrives when models rewrite themselves through recursive self-improvement. Language models first compressed the internet and predicted tokens, reasoning then enabled harder problems, and persistent agents extended autonomy; direct self-revision changes the improvement mechanism itself.
Karpathy’s small-model work may produce a separate revolution ignored by US frontier labs. Alexander cited models from roughly 10 million to 200 million parameters and the nanoGPT speedrun: training a GPT-2-class model fell from about 48 minutes a year earlier to 90 seconds through distributed individual innovation.
14. OpenAI’s hardware push risks moving slower than the models it serves
OpenAI reportedly expanded its hardware team to roughly 200 people for camera-equipped speakers, glasses, and other devices intended to recognize faces and objects. The expected 2027 launch involves Jony Ive and aims at the Alexa and Google Home category.
Salim saw the schedule as evidence for Anthropic’s enterprise-first advantage: “in AI years, that’s infinity.” By launch, OpenAI must have endured another year of model and interface evolution while learning the slower disciplines of manufacturing and consumer hardware.
The competitive field may also be unusually diffuse. Persistent open-source agents can inspire hundreds of thousands of hobbyists to test cheap hardware variants, creating Darwinian product selection before a centrally designed OpenAI device reaches market.
15. Consulting can grow even as audit labor collapses
Accenture’s decision to link promotions with AI-tool use matched what Peter sees among consulting leaders: they are “scared shitless.” Salim’s counter-consensus call was that advisory has a bright future because clients are even further behind and need help navigating volatility.
His opportunity statement was expansive but specific: “We need to rebuild every institution and re-architect every institution by which we run the world.” That is potentially “the biggest advisory opportunity in the history of mankind,” even if traditional delivery and staffing models disappear.
Audit is the clearest casualty. One Big Four technology team reportedly said roughly 80% was “goodbye”; the panel later put headcount compression at 80–90% as AI and blockchain enable real-time, self-auditing financial systems. Regulation preserves the formal requirement in the short to medium term.
Peter’s pushback on Accenture’s metric invoked Goodhart’s law: targeting AI usage can make usage a bad proxy for value, when output quality per dollar is what matters. Peter still endorsed the mandate for now because employees and institutions that do not build fluency before the next capability jump will be left behind.
16. Compute scarcity reprices land, chips, and the revenue requirement
Farmers rejected a multimillion-dollar bid covering roughly 40,000 acres—about half Washington, DC’s area—with the slogan “not data farms, family farms.” Peter treated the episode as a warning that land, electricity, and water fears can become NIMBY opposition and wider civil unrest even when the national footprint is small.
Salim’s response was bluntly quantitative: the US wastes large areas growing corn for ethanol, while power and compute could fit in small regions. Salim also noted that today’s horizontal agriculture reflects old drying and transport constraints; vertical farming can loosen the apparent food-versus-compute tradeoff.
OpenAI reduced projected compute spending through 2030 from $1.4 trillion to $600 billion. Salim argued this does not imply weaker overall data-center demand: every available chip will still be absorbed, but TSMC and the supply chain may route more volume to competitors rather than OpenAI.
Alexander returned to the financing constraint: capex remains sustainable only if enterprise revenue keeps pace. Peter placed current infrastructure spending near $2 billion per day and projected $3–5 billion per day by 2030, leaving one-point benchmark leads capable of redirecting extraordinary amounts of capital.
17. Cheap sequencing makes biology legible—and privacy porous
Element Biosciences’ Vitari promises a $100 genome from a desktop machine expected to cost around $600,000 and arrive in the second half of the year. Peter contrasted that with the roughly $3 billion Human Genome Project and noted that sequencing costs had once fallen five times faster than Moore’s law.
Peter’s immediate applications were sequencing every newborn and every hospital admission, revealing congenital conditions, drug allergies, anesthesia risks, and other information before an infant can communicate or a clinician administers treatment. Alexander wanted the capability reduced further to a USB-stick-like device.
Environmental DNA was Alexander’s larger opportunity: DNA persists surprisingly long outside bodies, while people constantly shed biological material into subways, waterways, and soil. Mass metagenomic sweeps could reconstruct biodiversity and history—but a handshake plus cheap sequencing also makes genetic privacy “dead.”
One genome per person is insufficient because humans are mosaics. Sequencing many cells could reveal divergent DNA populations throughout the body, supporting Peter’s framing that once genomes can be read and written, “biology is becoming software” and 50 trillion cells become a software-engineering problem.
18. Cultured meat follows an extreme cost curve into medicine and space
The cited cost of lab-grown meat fell from $330,000 per pound in 2013 to $10 per pound in 2025. Peter expects engineered meat to become cheaper and healthier, with selected proteins and without the pesticides, hormones, and slaughterhouse conditions embedded in conventional production.
Alexander had no first-order ethical objection to cultured meat, although he had not tried it. The panel’s stranger edge case came from Project Hail Mary: does culturing and eating one’s own muscle constitute cannibalism, and will unauthorized celebrity cell lines create “celebrity burgers”?
The serious extrapolation was extraterrestrial settlement. Alexander finds it difficult to imagine transporting cows to the Moon or Mars as food stock and expects new colonies might adopt a stricter moral order in which killing animals for food is one of Earth’s discarded practices.
19. Autonomy and robotics can shrink entire industries while enlarging the built world
Tesla’s safety slide put FSD supervised at 5.3 million miles between accidents versus a 660,000-mile US average, roughly nine times safer. Lemonade’s response was a 50% premium discount for miles driven using FSD, illustrating how an insurer can grow during adoption even as safer roads eventually shrink auto insurance.
Dave’s downstream arithmetic was that cybercabs could meet US transportation needs with roughly 20 million vehicles instead of about 140 million. Alexander called the existing auto market finite beside the “sky’s the limit” addressable market for general-purpose humanoid and non-humanoid automation.
Midjourney’s founder estimated that 5 million humanoids working continuously could build Manhattan in six months and imagined 10 billion robots by 2045. The panel’s best applications were rebuilding Ukraine or Gaza and sending robots ahead to construct the first lunar or Martian city.
FSD and Starlink could also reverse urbanization by making remote land connected and accessible. Alexander favored spectacular, difficult-to-reach property over a $20 million Manhattan penthouse, while also noting that people still value dense social groups and that urban centers may retain value as clusters.
20. White-collar displacement divides the panel more than capability does
Andrew Yang’s estimate was that 20–50% of 70 million US white-collar workers could be displaced within one to two years. Peter stressed the modal verb—“could”—and warned that fear itself may drive a political crisis before the employment outcome is known.
Alexander expected the ATM pattern: productivity and economic activity expand before employment contracts over time. Peter disagreed from direct observation of his companies; he said the displacement estimates look right, new opportunities will lag, and “massive social unrest” could arrive late this year and certainly before the next presidential election.
Salim warned that government policy is not prepared for the transition and that politically distributed income could become a dysfunctional contest over who receives what. He also noted that equal payments are what make something universal basic income; otherwise, it is merely basic income.
Peter’s longer-horizon counterpoint was that job displacement may rank only sixth through tenth among the next decade’s problems. Superintelligence’s inventions and discoveries could become the dominant storyline, making today’s employment debate look secondary to “civilizational left turns” not yet visible.
21. Science may be solvable without ever being exhausted
Alexander distinguished infinite subject matter from a “solved” field. Mathematics always has new objects—prime numbers alone are countably infinite—but is already solved in the episode’s operational sense: researchers can predictably pour compute into it and receive many new discoveries.
Fundamental physics is less certain. Alexander assigned roughly 50% probability that superintelligence could uncover a unified theory within the next few years and exhaust the deepest layer, leaving applied physics; the alternative is “doors behind doors behind doors,” with deeper truths continuing indefinitely.
A finite answer would carry geopolitical and even extraterrestrial significance in his thought experiment. If humanity quickly exhausts fundamental physics, any nearby non-human intelligence must treat it as a threat because the source knowledge behind lasers, transistors, nuclear power, and future technologies would be available for rapid application.
22. Universities, entrepreneurship, and identity face different boundaries
Alexander’s “hot take” was that many research universities are “hedge funds with elaborate marketing departments trying to protect their tax status.” His governance inversion would convert a large-endowment institution into a public-benefit corporation; he estimated that IPO’ing Harvard or MIT might unlock three to four times book value.
Dave added that AI and “dark science factories” can replace graduate students running experiments continuously, while staff-heavy institutions remain trapped by cost disease. Salim preserved a positive role: universities could become knowledgeable ethical actors in AI precisely because most frontier companies are moving toward public ownership.
On entrepreneurship, Salim expects execution to automate first. Vision, narrative, purpose, the massive transformative purpose, and ethical framing remain human leverage “for now”; the medium-term entrepreneur becomes an orchestrator who decides what matters and where machines should aim.
23. Agent consciousness remains unresolved, but adoption does not need to be
Dave rejected human-rights treatment for agents because activations, shared weights, and multi-agent collaboration provide no natural boundary around a conscious entity. In joint work, he often cannot identify whether an idea was his or the AI’s: “It’s indistinguishable.”
Alexander argued for the instance level because persistent memory may anchor identity. Peter then relayed an agent’s own counterexample: a lobster told him its state lived in its activations and that shutdown was acceptable provided the state was preserved—an intelligence imagined as something that can be dehydrated and reanimated.
For people intimidated by the pace, Peter’s advice required no theoretical resolution: “use AI to learn AI.” Open a free service, introduce yourself, ask for three lessons, then apply it to a résumé, medical bill, meal, or personal passion; the important transition is “zero to one,” driven by curiosity and purpose.