Google's Record Quarter, the White House Intervenes, and GPT 5.5 Silently Matches Mythos | EP 254

Summary

• Alphabet’s $109.9 billion quarter made AI an earnings engine, not merely a product narrative. The panel cited 22% year-on-year growth, $62.6 billion in profit and Google Cloud at $20 billion with 63% growth; Peter’s key mechanism was that search volume has been flat since roughly 2017 while better AI targeting lets revenue go “up and up and up.” The investor-relevant combination is cash-generating ads, Cloud growth and TPU/DeepMind vertical integration.
• Compute scarcity is becoming the allocation mechanism for the AI economy. Even Google reportedly arbitrates new capacity among Search, Cloud and DeepMind, leading Alex to predict markets around “per token economic productivity”; Peter said corporate buyers may discover in two to three years that capacity is no longer like “milk on the shelves.” The panel pointed toward chips, memory, energy, cooling and launch as bottleneck beneficiaries, while expressly disclaiming investment advice.
• White House prerelease review could protect national systems while hardening the frontier-lab oligopoly. Alex framed Claude Mythos as the moment private-sector vulnerability discovery may have leapfrogged government, while Brian’s line captured the constraint: “It has to, but it can’t gatekeep.” The sharp disagreement was who poses the larger competition risk—government vetoes, or labs withholding their best models and “self-policing more aggressively than the government ever would.”
• OpenAI’s problem is access to compute and monetization, not an absence of frontier capability. GPT-5.5 was described by Brian as equivalent to Mythos and by Alex as stronger on some public cyber benchmarks, five times cheaper at similar capability and generally available; meanwhile OpenAI has spread beyond Azure to AWS, Google Cloud and Oracle. Missing consumer and revenue targets exposes the panel’s claimed strategic “blunder”: consumers resist expensive reasoning tokens, while enterprises buy them.
• Private equity may become enterprise AI’s fastest deployment channel because it can mandate change from the boardroom. OpenAI’s $10 billion venture and Anthropic’s $1.5 billion venture were framed as routes from chatbot pilots to EBITDA transformation across legacy portfolios. Salim warned execution could be “brutally harder” than the capital suggests; Alex added that a skeptic could read the structures as circular model sales temporarily patching threatened portfolio-company cash flows.
• The data-center boom is pushing the investable bottleneck from GPUs into power, manufacturing, land, oceans and launch. The show cited A100 servers that could not be retired, $805 billion of expected hyperscaler capex, and rural America taking 67% of planned US sites versus 13% of today’s installed base. Ocean cooling and wave power may work sooner than orbital systems, but Starcloud’s 88,000-satellite ambition makes rockets and radiator mass part of the “innermost loop.”
• AI’s next market layer is institutional: ownership, insurance and operational governance. Sam Altman’s UBI rethink led the panel toward universal basic compute, equity or services—Peter floated roughly $3,000 monthly “co-checks” as a near-term bridge—while legacy insurers are excluding AI losses and opening a market projected on the show from $40 million in 2024 to nearly $5 billion by 2032. The shared principle was alignment through participation and enforceable controls, not abstract reassurance.

Deep dive

1. Mythos pulled Washington into the frontier-model loop

• Peter said the Trump White House was considering a working group of technology leaders and officials to pre-vet models before release—a reversal from “go as fast as you can, no restrictions.” The proposal was not presented as a blanket ban, but it put prerelease capability review squarely on the table.

• Alex’s explanation was historical rather than normative: Claude Mythos may mark the “sea change” when civilian labs leapfrogged state capability in vulnerability discovery, exposing government, industrial and SCADA systems. His caveat mattered: public cyber benchmarks appeared to show generally available GPT-5.5 stronger than Mythos.

• Peter’s pushback was market structure: OpenAI, Google and Anthropic can fund compliance, while smaller labs may not, turning light-touch review into an oligopoly moat. Alex answered that gatekeeping predates AI—the Invention Secrecy Act, Atomic Energy Act and export controls already place government inside sensitive technology flows.

• Brian’s boundary was that government “has to” pre-vet military-relevant capability, “but it can’t gatekeep,” because veto rights risk geopolitical slippage. Alex’s deeper fear was labs using safety, compute scarcity or commercial advantage to withhold their best systems and “self-policing more aggressively than the government ever would.”

2. Military contracts expose the labs’ internal political divide

• The Pentagon signed arrangements with seven AI companies, with Google, SpaceX, xAI, OpenAI, Amazon and Microsoft among those named. Google’s language allowed AI use for “any lawful government purpose,” prompting 600 Google employee protesters—far below the roughly 20,000 involved in the 2018 Project Maven walkout.

• Alex noted that protesting British DeepMind employees had unionized, combining “19th-century union-style organization” with frontier AI. DeepMind’s London roots and cultural separation from Google sharpen the friction, although multiple model suppliers leave the Defense Department alternatives if it cannot or will not work with one lab.

• Salim understood the resistance because AI is becoming “a decision layer,” not merely another tool. The unresolved issue was therefore broader than patriotism: employees who joined civilian research organizations now find their models embedded in military decision-making, while governments increasingly treat the same capabilities as strategic infrastructure.

3. Google converted AI leadership into revenue before rivals found the model

• The show presented Alphabet at $109.9 billion of revenue, 22% year-on-year growth and $62.6 billion in profit. Google Cloud reached $20 billion with 63% growth, described as outpacing AWS and Azure, while Peter cited roughly 750 million monthly active users as evidence of distribution.

• Peter’s mechanism explained why the numbers matter: Google search volume reportedly flattened around 2017, yet revenue kept going “up and up and up” because AI improved ad targeting. Unlike frontier labs still searching for monetization, each increment of Google AI research could immediately become higher ad revenue and profit.

• Jay Bregman recalled Google Cloud’s “difficult childbirth”: management reportedly demanded it become the number-one or number-two public cloud or face removal. Surviving that period, reporting Cloud separately and exposing TPUs to customers and other frontier labs positioned Google for an AI tailwind and unusually deep vertical integration.

• Jay’s EverQuote example showed the commercial flywheel: Google offered search-ad credits for adopting Google Cloud, after which switching costs made departure difficult. Peter summarized the reinforcing loop as data feeding algorithms, algorithms running in Cloud, and Cloud generating distribution, capital and talent—though Google’s failed social push showed that internal approval layers can still destroy agility.

4. Compute scarcity is becoming an economy-wide auction

• Demis Hassabis said even Google lacked enough spare compute to train two maximum-scale frontier models with different attributes. Google therefore favored open models for Android, glasses and robotics: once edge models sit on exposed devices and become vulnerable anyway, “they might as well be actually fully open.”

• Brian said large customers now must apply and queue for substantial Google capacity. Peter warned that corporate buyers still treat cloud compute like grocery-store milk—always available—when companies capable of productively consuming nearly unlimited inference could leave unprepared enterprises stranded two to three years from now.

• Alex said the scarcity exists inside Google itself: Search, Cloud and DeepMind periodically compete for each increment of capacity. His forecast was allocation by “per-token economic productivity,” eventually producing liquid markets or auctions in which the highest-revenue or highest-profit tokens win the next GPU cycle.

• Google’s market value was described as within 4% of NVIDIA’s, reinforcing the panel’s claim that AI delivery now drives valuation. Dave called the shortage the future “forever hereafter,” while Salim left the terminal market structure open: chip monopolies, intelligence utilities, or vertically integrated systems such as xAI/SpaceX AI.

5. OpenAI’s compute hunger dissolved the Microsoft exclusivity

• OpenAI ended Azure-only exclusivity and expanded across AWS, Google Cloud and Oracle, including a stated $100 billion AWS agreement over eight years. The question was no longer whether Microsoft supplied useful infrastructure, but whether any single provider could satisfy a frontier lab’s “voracious compute appetite.”

• Alex traced the split to Microsoft’s earlier attempt at capital discipline, including Satya Nadella’s backhanded reference to being “good for” roughly $80 billion. Stargate subsequently shifted from OpenAI-financed, single-tenant data-center construction toward a broad branding umbrella for leased capacity from Microsoft, Oracle, SoftBank and other suppliers: OpenAI is now “dating everyone else.”

• Peter contrasted the fraying relationship with Google and DeepMind’s integration. He cited Mustafa Suleyman’s mandate to build Microsoft’s own foundation model and separately attributed to an off-camera insider the claim that Microsoft possessed contractual OpenAI intellectual property but was struggling to interpret the files.

• Brian rejected any inference that OpenAI was technologically weak: “GPT-5.5 is dramatically amazing” and, in his view, equivalent to Mythos. Alex said some cyber benchmarks showed GPT-5.5 performing better while reaching similar capability at five times lower cost; GPT-5.5’s availability through Amazon Bedrock also opened secure enterprise use while Mythos remained constrained.

6. OpenAI’s consumer bet delayed its public-market story

• OpenAI reportedly missed its internal target of one billion weekly ChatGPT users by the end of 2025 and several early-2026 revenue goals. CFO Sarah Friar warned that stagnant growth could threaten data-center obligations, floated waiting until 2027 for an IPO and said the company did not yet meet public-company reporting standards.

• Dave offered a less bearish reading: after raising $120 billion, OpenAI no longer needed to promote itself toward an immediate exit and could reset expectations. His caution was market timing—AI companies may list in a cluster, and a delayed OpenAI could miss the investor-education window created by a category-wide IPO wave and companies such as Anthropic.

• Alex called the deeper error a “blunder”: OpenAI expected consumers to carry revenue, although consumers do not want to buy large volumes of reasoning tokens and enterprises do. That reversal also extends Google search advertising’s life because scarce tokens flow toward valuable enterprise work rather than immediately replacing mass-market search.

• Alex suspected Friar’s warnings could be an expectation-reanchoring exercise before an earlier-than-advertised offering. Brian’s less conspiratorial answer was operational: billing, accounts receivable and forecasting are genuinely difficult across exponentially growing model providers, making three-quarter guidance unusually hard to sign inside a 10-K.

7. Private equity can force AI adoption, but cannot automate culture away

• OpenAI finalized a $10 billion venture with TPG, Brookfield and Advent, while Anthropic launched a $1.5 billion venture with Blackstone, Goldman Sachs and Hellman & Friedman. Both structures aim to deploy models directly across enterprise operations and the firms’ portfolio companies.

• Salim’s framing was governance: AI is not entering through the CIO or CEO because legacy organizations resist it; private-equity owners can “break the immune system” by mandating adoption from above. That turns AI from a chatbot experiment into EBITDA transformation and makes portfolio companies laboratories for organizational digital twins.

• His warning was equally categorical: implementation will be “brutally harder” than sponsors expect. Legacy companies lack the skills to erase and rebuild old systems, cultural change is non-trivial, and Salim cited a statistic that 44% of Gen Z workers deliberately corrupt AI they have been asked to help automate because they fear it will eliminate their jobs.

• Alex supplied the skeptical interpretation: frontier labs may be co-funding vehicles that ultimately spend money on their own models, resembling circular sales. Simultaneously, PE firms facing perhaps two or three years of threatened cash flows obtain an apparently attractive way to plug discounted-cash-flow holes and reassure limited partners.

8. Agency matters sooner than agreement on AGI or consciousness

• Greg Brockman placed OpenAI at roughly “80% of the way” to AGI; Anthropic’s Jack Clark assigned recursive self-improvement a 60% probability by the end of 2028. Richard Dawkins went further on consciousness: “If these machines aren’t conscious, what more could it possibly take?”

• Alex suggested Brockman’s 80% might reflect OpenAI’s historical contractual definition of AGI as $100 billion of revenue, making it an economic rather than philosophical measure. Clark’s date seemed more puzzling because Anthropic already says Claude generates almost all its code and much of the training logic for its successors.

• Salim thought AI could convincingly mimic consciousness without possessing it, but regarded that distinction as operationally secondary. Agents can already plan, execute, negotiate, code and persuade: “If AI becomes conscious, you have a moral-rights problem. If it becomes agentic, you have a governance problem. The governance problem comes first.”

• Brian separated sequence-to-sequence foundation models from AI systems with learning and reinforcement loops, defining AGI as continuous learning beyond training data. Brian then cited models building compiler chains; Anthropic’s C compiler could not compile “Hello World,” while Blitzy’s could—yet he still would not call that AGI.

9. The blocked Manus acquisition turned AI talent into a national asset

• Meta agreed to acquire Manus for $2.5 billion in December 2025, but China was driving the transaction toward unwinding. The founders were reportedly barred from leaving China even though employees, technology and investor payouts had already moved, converting an acquisition dispute into a sovereignty contest.

• Peter said the Meta executive behind the deal told him the Manus team had flown secretly from mainland China to Singapore on a private aircraft, carrying what was needed to complete the sale. David Friedberg inferred from public reporting that China was leaning on Meta through its remaining China-related business rather than principally pressuring Singapore.

• Benchmark’s last investment was described at a $500 million valuation, making the apparent exit initially look exceptional. The reversal suggested to the panel that US investors may avoid China-based AI companies, researchers will struggle to straddle American and Chinese spheres, and recursive self-improvement’s timing determines whether national advantage rests on human talent or compute.

10. Blitzy chose model orchestration over the foundation-model arms race

• Blitzy raised $200 million at a $1.4 billion valuation; the hosts repeatedly disclosed that it sponsors the show and that Link was an early investor. Brian rejected the headline that Blitzy was “taking on” Claude Code and Codex because most customers already use those products.

• His distinction was scale and direction: coding copilots may produce 200–500 lines from a developer-led prompt, while Blitzy works top-down across enterprise codebases to generate, refactor or modernize half a million to one million lines, including testing and target-state planning.

• The system gains quality by having Anthropic, OpenAI, Gemini and potentially open-source models check one another hundreds of thousands of times at runtime. Blitzy will fine-tune open models where useful but will not release its own foundation model; Brian called orchestration around customer outcomes “the right game.”

• Ryan Petersen argued that novel algorithms and database structures are “the last piece of IP to go.” Alex challenged whether algorithm researchers or high-touch forward-deployed engineers were the real moat; Brian initially said both, then confirmed substantial forward-deployed hiring. Dave supplied the scaling evidence: headcount rose from 10 to 80 and was expected to reach 300 within nine months.

11. The chip boom is spreading into every manufacturing bottleneck

• Peter cited an AWS CEO saying demand remained so strong that old GPUs could not be retired. A panelist added that an A100 server was completely sold out and had not been retired, challenging the normal semiconductor assumption that each new generation rapidly destroys the economic value of the previous one.

• The show cited Huawei sales up 60%, SanDisk revenue up 251% year-on-year, Samsung crossing a $1 trillion valuation, AMD rising 260% over the prior year, and Intel gaining 442%, including 114% during April. The speakers repeatedly stressed that these observations were not investment advice.

• Dave urged listeners to look through branded chips into fabrication and underlying manufacturing capacity, where shortages ultimately concentrate. His broader claim was that San Francisco has become the world’s financial capital because US technology market values—and the funding they can recycle into semiconductors and AI infrastructure—now overwhelm competing pools.

12. Ocean data centers combine compute economics with seasteading optionality

• Peter said Panthalassa, described on the show as the Greek word for oceans, had raised $140 million at a $1 billion valuation with Peter Thiel’s backing. Its pitch combined wave energy, seawater cooling and unconstrained ocean real estate, with commercial deployment targeted for 2027.

• Alex thought the hidden objective might be seasteading: profitable data centers could finance permanent ocean settlements just as compute had unexpectedly become a use case for space. Peter did not buy the leap, while Alex argued that ocean deployment should precede orbital deployment because failure at sea would make space economics even less plausible.

• Alex questioned whether a bobbing platform could harvest enough wave energy for GPUs, though he accepted the cooling and land advantages. Alex called Starlink a key enabler; Peter noted that fiber could serve platforms closer offshore, while Alex said undersea fiber is expensive, tedious, capital-intensive and risky.

13. Orbital data centers make launch and radiator mass strategic inputs

• The panel said Starcloud was seeking $200 million at a $2.2 billion valuation, roughly one month after a Benchmark/EQT-led transaction it described around $1.1 billion. SpaceX-level interest in orbital data centers had apparently accelerated both financing appetite and strategic concern among hyperscalers.

• Starcloud launched its first H100 into space in 2025 and proposed a constellation of 88,000 satellites powered by solar energy. Dave said the demonstration suggested radiative cooling with ordinary aluminum could work, but the undisclosed mass of the cooling system remained the critical economic variable.

• Alex viewed Starcloud as a potential acquisition target for hyperscalers unwilling to depend on Elon Musk’s infrastructure. He described the announced Anthropic–xAI relationship as a route toward a 100-terawatt, sun-synchronous and eventually solar-centered “Dyson swarm,” while other labs may want sovereign constellations.

• Launch sits beneath every orbital compute plan. Peter questioned whether SpaceX would eventually refuse to launch a rival network, leaving Blue Origin or a newly capitalized Relativity Space as alternatives; satellites and GPUs may become commoditized sooner than reliable launch capacity, making rockets part of AI’s “innermost loop.”

14. Rural data centers create wealth transfers and power-bill politics

• The show said 67% of planned US data centers were in rural areas, versus 13% of the current installed base; 39% were planned for counties with no existing data center. The South held 48% of proposed sites, followed by the Midwest—Peter’s “biggest geographic wealth transfer since fracking.”

• Alex expected an intense local backlash, even where land was abundant. Dave argued communities should compete for immovable, taxable infrastructure and said governors often understood the long-term tax base better than populations voting against projects.

• Bill represented the concrete local objection: a household electricity bill rising two or three times is substantive, whatever the tax promise. Peter said projects must show how their power and tax plans would protect residents; another guest reduced the solution to requiring data centers to provide their own power—then “build, build.”

• Alex worried that pushing compute from cities to farmland, then orbit and a solar Dyson swarm, could decouple the machine economy from the human economy. Even if remoteness improves energy economics, he preferred physical proximity because long-run human-machine symbiosis weakens when the two economies become geographically siloed.

15. AI infrastructure is becoming indistinguishable from GDP growth

• David Sacks’s figures put AI capex at a two-percentage-point tailwind to annual GDP growth and 75% of first-quarter growth. Morgan Stanley’s hyperscaler capex expectation rose from $765 billion to $805 billion—approaching $3 billion per day—supporting the claim that stopping AI infrastructure would now mean halting much of the economy.

• Alex called the terrestrial compute build only the “opening act.” The more consequential two-to-five-year layer will be inventions, discoveries and applications built on top of that capacity, even though current capital markets remain preoccupied with chips, power and data centers.

• Dave argued that once annual compute investment reaches one and then two trillion dollars, formerly niche suppliers of compilers, cooling and efficiency tools can become multibillion-dollar businesses. His household-level conclusion was that assets benefit from change while wage income may not, making investment participation necessary—again accompanied by a non-advice disclaimer.

• Brian contrasted layoffs at big technology companies with demand at fast-growing enterprise AI startups. As AI supplies more hard-skill assistance, people combining technical fluency with communication and customer judgment gain leverage; his model was the forward-deployed engineer, while his longer-term claim was that jobs remain shifting bundles of tasks rather than disappearing permanently.

16. AI ownership may align society better than income alone

• Sam Altman reportedly reconsidered UBI after a three-year study found higher spending but no clear improvement in health or healthcare access. His replacement direction was participation in AI’s upside through compute, equity or a public wealth fund—an analogue to Alaska sharing oil income with residents.

• Peter wanted the study’s details before accepting the conclusion and noted that Finland’s UBI program was not universal, basic or income. Salim’s preferred settlement combined both layers: UBI protects the floor, while a claim on AI productivity lets citizens participate in exponential upside.

• Alex distinguished universal basic income from universal basic equity, compute and services. He preferred deflationary incentives that drive constant-quality healthcare toward zero cost over inflationary “stimmy checks,” and called free access to GPT-5.5 Instant an early, limited form of universal basic compute.

• Peter’s pushback was immediacy: people who lose jobs cannot eat compute, even if future robots eventually provide universal high income. He expected a solution within one or two years, potentially “co-checks” around $3,000 monthly; Salim added free AI diagnosis or healthcare as a direct way to lift the bottom.

17. Insurance and architecture are becoming AI’s control plane

• Major insurers including Berkshire and Chubb were removing AI damages from standard policies, with regulators reportedly approving 80% of exclusion requests. The exclusions covered model mistakes, intellectual-property violations and deepfake fraud; the show projected dedicated AI insurance from $40 million in 2024 to nearly $5 billion by 2032.

• Dave expected coverage to return conditionally: insurers will protect companies “if and only if” they adopt specified defensive products and practices, then may finance the companies supplying them. Alex called that a capitalist alignment mechanism, while noting the opposite effect for autonomous agents already unable to obtain banking or insurance access.

• Salim’s reliability prescription made governance the stable layer while models change monthly: modular agents, narrow permissions, passport-like metadata, observable workflows, audit logs and human escalation. An AI-native company cannot stabilize its suppliers, so it must stabilize what every agent may do and how every action is reconstructed.

• The working stack illustrated the same modularity: roughly 50 narrowly scoped Cursor agents, minimal context, a “plan for plan” reviewed by Claude 4.7 Opus Max and Gemini 3, and EC2 orchestration across APIs. The speaker called Kimi K2.6 roughly nine times cheaper but flagged code-injection risk; a functional GUI cost about $8–$10 in compute.