Claude Code Ends SaaS, the Gemini + Siri Partnership, and Math Finally Solves AI | #224
Summary
Claude Code with Opus 4.5 is the episode’s clearest near-term discontinuity: longer autonomy horizons are turning software from artisanal craft into an industrial process. Alex Wissner-Gross sees autonomy potentially growing hyper-exponentially from hours toward weeks, months, and years; Peter Diamandis says running five to 10 concurrent agents now produces code faster than he can mentally track, with his recent output exceeding everything he wrote previously. “Only the paranoid survive” is the investable warning: incumbents have access to the same frontier models, but static products and slow management teams can be repriced abruptly.
NVIDIA is expanding from chips into an integrated operating stack for physical reality. Cosmos creates physically plausible training environments, Alpamayo connects camera input to vehicle action, and Vera Rubin combines CPU, GPU, memory, interconnect, and housing around the data center as computing’s new form factor. Peter calls this NVIDIA becoming “the AWS of reality,” while Alex calls it classic “commodifying its complement.”
AI threatens both SaaS seats and knowledge-work headcount, but the panel rejects a simple incumbents-die narrative. McKinsey already describes itself as 40,000 humans plus 20,000 agents, up from 3,000 agents roughly 18 months earlier, while Salim Ismail thinks one agent per employee is “ridiculous” and expects nearer 100. The winners may be adaptable incumbents using the same frontier models, AI-native enterprise stacks operating beside legacy systems, and eventually “single-person unicorns.”
The trillions committed to AI infrastructure still require a monetization bridge that has not been proved. OpenAI’s cited compute rose from 0.2 GW in 2023 to 1.9 GW in 2025 while revenue climbed from $2 billion to $20 billion. Dave Blundin calls the parallel “correlation, not causation.” Alex’s “elephant in the room” is that consumers resisted costly, slower default reasoning and enterprises have not yet generated outcomes sufficient to support sustained tripling; transformative reasoning applications must arrive for the capex cycle to continue.
Google’s Apple distribution deal strengthens the most complete AI stack, without necessarily killing the web. Gemini-powered Siri and Universal Commerce Protocol could move transactions from search toward a “magic box that gives action,” yet Alex stresses that UCP standardizes conversational commerce rather than replacing browsing, websites, or independent shopping agents. Dave predicts Google will exceed NVIDIA’s market capitalization “by the end of next year,” while Peter expects frontier-lab capital requirements to provoke acquisition attempts even as Alex sees regulation blocking a broad recombination.
AI solving open mathematics is presented as the leading edge of a much larger scientific automation cycle. GPT-5.2 Pro, paired with formalization and verification tools such as Harmonic’s Aristotle, was reportedly solving notable Erdős problems several times a week because mathematics offers enumerable questions and clean evaluation. “Problems wait to be prompted” becomes Peter’s company-building thesis: assemble the data, tests, approvals, or guardrails that let models attack chemistry, physics, materials, surgery, biology, and medicine.
Energy—not merely chips—is becoming the binding strategic input to intelligence. China was cited at roughly 10,000 TWh of generation versus a flat 4,000 TWh for the US, with output 40% above the US and EU combined and solar generation rising 46% in 2024 and 48% in 2025. The panel’s disagreement is over execution, not urgency: supply-chain dependence and regulation constrain the US, while insufficient power risks leaving superintelligence and its economic gains underbuilt.
Deep dive
1. The humanoid boom will consolidate before it standardizes
Peter’s defining CES observation was the “physical manifestation of AI”: 148,000 attendees, 4,000 exhibitors, 1,200 startups, roughly 38 humanoid-robot companies, and 12 robotic-hand manufacturers. After years in which AI could remain inside a screen, Salim said 2026 is when “you won’t be able to ignore it. It’s coming at you.”
The historical base rate argues against dozens of enduring platforms. Peter compared the field with 253 active US automakers in 1908, reduced to about 44 by 1929 as Ford, General Motors, and Chrysler consolidated the market; 278 early tire companies provide the equivalent analogy for today’s robot-hand suppliers.
Peter’s internet-boom analogy is worth keeping: abundant entrants did not invalidate the category, because Amazon later bought businesses such as Diapers.com and Pets.com. Alex’s narrower concern is that humanoids currently look unusually similar, suggesting eventual competition on price and AI among Figure, Optimus, 1X, Apollo, Digit, and perhaps only about a dozen viable designs.
Salim challenged the metaphor itself: a true Cambrian explosion should produce radically different body plans, not rows of near-identical humanoids. Dedicated hand companies may struggle when Brett Adcock, Elon Musk, and Bernt are vertically integrating, although Alex noted that “hands are hard” and a mature industry might eventually stratify horizontally enough to support component specialists.
2. NVIDIA is building the operating stack for physical reality
Cosmos takes thin simulator output or a single image and generates physically plausible video aligned across language, images, 3D, and action. The implication Peter pressed is that synthetic training data could erode the advantage Tesla accumulated from real-world driving footage by making large training corpora much cheaper to create.
Alex’s qualification was the “march of the nines”: safety-critical autonomy still needs extremely rare road events that ordinary video collection or simulation may miss. Yet Cosmos and Alpamayo serve NVIDIA’s business even without eliminating proprietary data moats, because optimized software encourages Chinese and unconventional OEMs to build competing autonomy systems on NVIDIA hardware.
Dave widened the aperture beyond driving. A generic world model does not automatically capture magnetic confinement in fusion, atom-wide chip wiring, nanoscale surgery, zero-gravity construction, radiation, or unfamiliar gravitational fields; each domain still needs specialized spatial data and tuned models. Peter’s synthesis is that NVIDIA is attempting to become “the AWS of reality.”
Vera Rubin makes that vertical ambition explicit: Vera is the CPU, Rubin the GPU, and six co-designed chips share data as one system. Alex sees CPU, GPU, memory, interconnect, and housing converging into the data center as “the new form factor of de facto computing,” while AI-driven DRAM shortages had already doubled the price of his son Jet’s gaming-computer build in six or seven months.
3. Compute demand is escaping the old semiconductor cycle
Dave rejected the comforting analogy to previous DRAM bubbles: “This is not going to come and go.” He expects high-performance memory and GPU demand to “go to infinity,” arguing that manufacturers remain too frightened of a cyclical downturn to expand at the exponential rate AI requires.
TSMC’s cautious fab construction therefore creates an opening for Elon Musk’s vertical strategy. Dave cited Musk’s minimum $16 billion Samsung arrangement, potentially reaching $40 billion, but suspects Musk is buying time while pursuing his own fabs—an unsettling prospect for suppliers expecting to serve him indefinitely.
Peter imagined 6G devices becoming interchangeable dumb terminals connected to cloud compute. Alex said latency and bandwidth, including broader Starlink availability, could support that architecture; Salim suggested that compute could sit at the edge of the 6G cloud, while Peter said local compute would remain useful when independent connectivity matters.
Asked for an upload date, Alex refused a single singularity moment: online writing already permits a low-fidelity reconstruction. He would be “very disappointed” if a nondestructive, ultra-high-fidelity brain scan were not possible within five to 10 years, while placing destructive Kurzweilian or Moravecian nanobot-style uploading roughly 10 to 20 years away.
4. Davos recognizes the AI shock but has few institutional answers
Dave described a one-year transformation at the World Economic Forum: buildings previously occupied by banks and consultancies had become AI venues, with “every billboard, every banner” carrying the theme. His event expected 270 speakers and, by his estimate, roughly $1 trillion of represented AI R&D, including leaders from frontier labs.
The geopolitical setting was less harmonious. An eagle-covered America House occupied the center of the promenade while Donald Trump’s arrival and the Greenland dispute raised tensions; Dave counted helicopters, drones, radar, and “3,000 people with machine guns” as the event’s new security metric.
Peter highlighted an announcement that OpenAI expected to show its first hardware device in the second half of the year after paying Jony Ive a stated $6.5 billion for the device effort. The unknown form factor matters because voice, wearables, and glasses could shift AI from a destination users visit into a continuously available interface.
Dave found political responses slow, reactive, and often framed around elections, despite broad recognition of imminent prosperity alongside severe unrest. Peter put the planning window at “one to three years maximum, more in the one-year time”; Daniel Schreiber of Lemonade had shared a proposal for implementing universal high income as one concrete contribution.
5. Agents preserve consulting demand while hollowing out its labor model
McKinsey CEO Bob Sternfels described a 60,000-member organization comprising 40,000 humans and 20,000 agents, versus only 3,000 agents roughly a year and a half earlier. He once expected one agent per employee by 2030; he now expects that ratio within 18 months.
Salim thinks large consultancies may perform well because volatile, slow-moving clients need help and “in the land of the blind, the one-eyed man is king.” His objection was the ratio: one agent per human is “ridiculous”; he expects around 100 agents per person and a transition from hourly work toward shared-value or outcome-based economics.
Alex found an accounting irony in counting agents as heads: including them in per-capita productivity could statistically conceal the intelligence-driven productivity boom. His less ironic destination is the zero-human company, where capital and software agents substitute directly for organizational labor.
The threat is that consulting clients themselves may not survive the shock. The corresponding opportunity, Salim argued, is “the biggest advisory opportunity in the history of mankind”: rebuilding the institutions through which society operates, provided consulting firms move beyond their old delivery model toward institutional redesign.
6. The job singularity favors entrepreneurs over credentialed employees
Robinhood CEO Vlad Tenev’s “job singularity” is a Cambrian explosion of job families, micro-corporations, solo institutions, and single-person unicorns. The internet gave individuals worldwide reach; AI now gives them “a world-class staff,” turning entrepreneurship into a plausible default occupation rather than a specialist path.
Salim translated that into “future shock to future shape.” His workshops with teenagers assume that whatever employment means when students leave college in five or six years will differ radically from today; the practical counsel is to become “the entrepreneur, not the employee,” and a creator rather than a consumer.
Peter questioned whether college becomes “the absolute wrong move” unless used to find purpose, collaborators, or a company. Salim’s decade-old prediction that his son would not attend university now looks plausible—certainly not for the purpose of securing a conventional job—because higher education is structurally unprepared for the transition.
The uncertainty remained explicit: neither speaker offered a settled replacement for university’s developmental and social functions. Salim joked that parents still need somewhere to send their children, prompting Peter’s description of “adult daycare” as a possible job of the future.
7. Opus 4.5 turns coding autonomy into an industrial process
Peter cited the sharpest formulation: Claude Code with Opus 4.5 moves software creation “from an artisanal craftsman activity to a true industrial process”—comparable to the Gutenberg press, sewing machine, or camera. Alex called the combination “Clopus” and treated its longer METR autonomy horizon as the genuinely important capability.
Alex rejected calling the moment AGI because some form of generality has arguably existed for roughly five and a half years. The inflection is autonomy: Claude Code plus Opus 4.5 and GPT-5.2 Codex can execute many sequential actions, with reports of complete Rust web browsers and functioning JavaScript engines built from scratch rather than over years.
His stronger hypothesis is hyper-exponential autonomy—“an exponential of an exponential”—potentially pushing usable work horizons from five hours toward weeks, months, and years. He preserved the caveat that every point on an exponential curve feels like a knee, but said the hyper-exponential forecast increasingly fits what he observes.
Peter described the transitional human cost. Running five to 10 Opus 4.5 agents concurrently is more mentally taxing than writing code slowly, because architectural decisions and outputs arrive faster than one person can track; his Claude bill runs roughly $100 to $1,000 daily, while his recent code output exceeds everything he wrote previously.
8. Code generation reprices SaaS, but access to models is not differentiation
Salim’s former Yahoo developer friends were “walking around with their jaws dropped open,” asking, “How do I get my head around this? This is unbelievable.” Alex sees Anthropic’s focus on programming as an implicit bet that code generation is the shortest route to recursive self-improvement and broad labor substitution.
Peter’s provocation was whether individuals can rebuild Salesforce, SAP, or Stripe from prompts, killing both SaaS and vibe-coding vendors. Dave pushed back that the result would not necessarily be the death of incumbent companies: many have already pivoted substantially, and the market will not simply disappear.
Salim’s broader warning is that “the future of the world belongs to flexible companies” that can pivot and improve constantly. Peter supplied the related “Only the paranoid survive” framing. Salim also noted that six of the Magnificent Seven are doing something fundamentally different from what first made them large, while Microsoft and Oracle shifted heavily toward cloud revenue.
Alex played the contrarian card: incumbents possess the same “weapons of mass superintelligence” as customers building replacements. Bespoke systems will cancel some $500,000 CRM contracts, but heavily customized Salesforce deployments already reflect pent-up customization demand; on a global basis, shared tools may simply produce “a new equilibrium—ho-hum, nothing to see here.”
9. An AI-native enterprise stack may emerge beside legacy systems
Salim’s rebuttal distinguished improving systems of record from bypassing them. He expects teams to “red-team” traditional enterprise stacks from the side, creating an AI-native operating layer that does not depend on the legacy record architecture and could become visibly separate within roughly six months.
Dave reconciled the macro and micro views: markets always settle into equilibrium, but investors can lose heavily in the transition. Equal access to frontier models does not imply equal execution, so market capitalizations should shuffle between “lazy laggards” and management teams capable of sustained reinvention.
Talent movement becomes a leading signal under that framework. Dave said quantitative funds are already analyzing flows of people into and out of companies because leadership and incoming technical talent indicate whether broadly available AI capabilities will actually be used well.
The episode’s SaaS call is therefore conditional, not a blanket tombstone: narrow product classes and old operating models are exposed, while heavily customized systems and fast-moving leadership may survive. What disappears fastest is the assumption that recurring revenue itself constitutes a durable moat.
10. Gemini-powered Siri makes commerce agentic without extinguishing the web
Peter framed the Google–Apple partnership as a change in interface physics: users move from a search box supplying information to a “magic box that gives action.” Universal Commerce Protocol could embed native AI checkout inside the agent experience, eliminating many URLs, passwords, pop-ups, app handoffs, and conventional website flows.
He extended the question beyond commerce: if wearables, AR glasses, voice, and listening become primary interfaces, keyboards might recede and even reading skills could weaken. “Who the hell is going to be typing next year?” captured the speed of his forecast, though he left the eventual OpenAI hardware form factor unresolved.
Alex’s pushback is worth keeping: UCP is a JavaScript-oriented standard for e-commerce inside agentic conversations, “that’s all it is.” People browse for products, use the web for far more than shopping, and deploy purchasing agents that may not use UCP; Amazon’s disputed “buy it with an AI agent” effort illustrates the competing paths.
Salim ultimately agreed with Alex, citing Yahoo Mail tests in which moving the Send button only a few pixels caused usage to collapse. Human interfaces are deeply habitual, making even technically superior redesigns slow to displace familiar patterns; Peter kept the disagreement open as a bet to revisit in several years.
11. AI capex needs transformative reasoning revenue to remain financeable
OpenAI CFO Sarah Friar’s charts paired compute growth—0.2 GW in 2023, 0.6 GW in 2024, and 1.9 GW in 2025—with revenue rising from $2 billion to $20 billion over 2023–2025. Peter interpreted the release as an investor argument that more data centers generate proportional economic value.
Salim read it as possible preparation for an IPO: unlike Meta, Google, or even xAI, OpenAI lacks an established “infinite cash-flow machine” and must finance chips, data centers, and energy externally. Dave rejected the causal claim, calling the matching lines convenient correlation shaped by other factors on both sides.
Alex’s “elephant in the room” was the trillions in capex that must eventually earn an enormous revenue pool. Ads can fund only part of it, while GPT-5 reasoning by default exposed consumer resistance to slower, costly answers—especially when many users prefer instant, agreeable responses to thoughtful, non-sycophantic ones.
Enterprises do consume reasoning, but Alex has not yet seen outcomes transformative enough to rationalize sustained tripling of revenue and compute. The “field of dreams” assumption—build compute and revenue will come—survives only if frontier labs produce applications valuable enough to make customers willingly consume much more expensive inference.
12. Google’s stack strengthens as the frontier shifts toward embodied AI
Peter cited Alphabet reaching a $4 trillion valuation after a 65% stock rise, with custom TPUs, models, interfaces, distribution, and now Siri compounding into one stack. Dave made the explicit call that Google would surpass NVIDIA’s market capitalization “by the end of next year.”
Alex narrowed the alleged frontier-lab field. Google DeepMind qualifies; Microsoft and Apple arguably do not currently operate frontier labs, Amazon emphasizes infrastructure and efficient smaller models, and Meta was rebuilding after Llama 4’s perceived failure. He expected OpenAI, Anthropic, and xAI to survive, while treating Tesla as a frontier VLA provider rather than a conventional chatbot lab.
The definition itself may change within three years. Once humanoids combine vision, language, and action, frontier status could mean deploying capable robots rather than leading agentic chat; Alex therefore expects OpenAI to offer humanoids and cited xAI’s Grok and Tesla’s FSD 14.2.2 as an early form of that convergence.
Peter predicted Google or Amazon might try to acquire Anthropic before its IPO; Salim would advise Anthropic to go public first. Alex saw antitrust review as the binding constraint and doubted broad consolidation, aside from a Tesla–xAI–SpaceX combination, while Dave questioned whether the OpenAI–Elon Musk trial would even start on time and noted the potentially $1 trillion stake.
13. Verifiable math is the beachhead for automated discovery
Alex said notable numbered Erdős problems were beginning to fall several times per week, often through GPT-5.2 Pro paired with a formalization and verification system such as Harmonic’s Aristotle. His forecast is a progression from a trickle into a flood of valuable open problems solved in bulk.
Mathematics goes first because its problems can be enumerated and answers checked cleanly, not because other disciplines are beyond the models’ intelligence. Alex expects the process to “walk out of math” into physics, chemistry, materials science, biology, medicine, and eventually the humanities.
Peter turned verification into a startup filter: identify a field where AI lacks data, evaluations, tests, regulatory approval, or guardrails, then supply the missing unlock. The company that does so for chemistry, surgery, or another constrained domain could become the next Mercor.
Alex seized on the line “Problems wait to be prompted,” because imagination may now be the bottleneck. He removed even that comfort: models can generate the prompts and decide which unknown questions deserve attention. Peter already has Gemini write prompts for Claude, though he still reviews their alignment.
14. Inference specialization opens a hardware flank against NVIDIA
OpenAI’s Cerebras partnership suggested to Dave that training and inference are decoupling, with perhaps 80%–90% of compute already devoted to inference. Cerebras’s wafer-scale chips run hot and are specialized, but their speed makes them attractive when long reasoning chains require hundreds or thousands of tool calls.
Alex’s technical instruction was “follow the money and follow the SRAM.” Cerebras and Groq with a Q—cited as acqui-hired by NVIDIA for $20 billion—avoid some DRAM dependence by placing fast SRAM close to compute, giving OpenAI diversified supply and much higher throughput before a possible IPO.
The constraint is model capacity: SRAM cannot hold arbitrarily large models. Dave nevertheless described a potential NVIDIA vulnerability if training can be refactored into small enough pieces for that architecture; because index funds and 401(k)s carry broad NVIDIA exposure, an abrupt breakthrough would reach far beyond specialist semiconductor portfolios.
xAI’s Colossus 3 supplies the other scale extreme: a stated 2 GW, $20 billion facility whose projected build time was described as faster than prior phases, with the clip suggesting fewer than 122 days. Its “Macrohard” thesis—roughly four employees per GPU—packages a virtual “digital Optimus” that could substitute for enterprise software and knowledge workers, though Alex called it the most legible capex story rather than the most imaginative.
15. Energy capacity determines who can scale intelligence
Peter’s numbers were stark: China generates about 10,000 TWh versus a largely flat 4,000 TWh in the US, produces 40% more electricity than the US and EU combined, and moved from sixth place in 1985 to first in 2024. Chinese solar generation increased 46% in 2024 and another 48% in 2025.
Salim described a bifurcation between countries possessing talent and those possessing energy. China’s control of solar-panel supply chains helps explain US resistance, while decades of financially engineered offshoring left America unable to respond as quickly as China did when denied advanced GPUs.
Alex argued that the US has repeatedly been “scared of energy”: nuclear after Three Mile Island, fossil fuels because of carbon, and solar because of import and infrastructure vulnerabilities. If superintelligence arrives near an AI 2027-like timetable, he believes powering it should outrank legacy fears whose harms unfold on much longer time scales.
The geopolitical alternative is Chinese infrastructure supplying both energy and inference abroad: 20 African countries reportedly imported 2 GW of Chinese solar panels in one month. Dave’s investment caveat was obsolescence risk, but even a rapid fusion breakthrough would not instantly solve shortages because generators and turbines are already sold out; Boom Supersonic’s generator pivot illustrated that bottleneck.
16. Post-work institutions need agency, accountable founders, and new liability law
Salim grounded human agency in dignity and technological demonetization: broadly available AI lets individuals become extraordinarily productive, while lagging institutions create the psychological shock. Alex expects capitalism to thrive initially because capital substitutes for labor, then yield to an unknown “economics 2.0” built around partial, rather than universal, post-scarcity.
Dave’s founder advantage over the next three to five years is empathy: anticipate what people will want in abundance and identify the data or components that unlock a new AI capability. As execution becomes cheaper, Salim says the founder shifts from great doer to guardian of vision, purpose, and culture; Alex adds the less glamorous role of “the neck to wring” when a one-person unicorn causes harm.
Peter guessed robotaxis could move from marginal adoption to more than 50% of cars on the road within three or four years once regulation permits. The catalyst is a resident “Jarvis” that anticipates schedules and places a Waymo or Cybercab at the door, removing even the friction of ordering a ride.
Liability remained unresolved: Alex placed likely training-time responsibility on the lab but expects new case law and perhaps AI personhood for autonomous inference-time conduct. Salim countered that autonomy avoids 99.99% of contrived trolley cases and urged, “Let’s automate first.” Salim contrasted the US labs’ contained approach, which makes responsibility clearer, with China’s roaming models that improve beyond any single controller.