Big Tech’s $3T AI Bet, Claude Writes 90% of Anthropic’s Code, AIs Editing AI Brains | EP #308
Summary
- Big Tech’s $3T AI commitment now faces a financing constraint alongside its growth thesis. Alphabet, Meta, Microsoft, Amazon and Oracle have added roughly $1T of commitments in three months, with some leases stretching 30 years. Blundin expects competition and recursive self-improvement to keep spending high, while tighter lending terms and hedging expose the timing risk. Ismail’s summary: “The directionality is right. Timing it is very, very difficult.”
- Insurance float could become the next major funding pool for AI infrastructure. Wissner-Gross points to approximately $6T in US life-insurance capital and $3T in property-and-casualty insurance. Blundin proposes exchanging investment for reserved compute capacity, giving insurers infrastructure exposure and resources for their own automation. The challenge is underwriting an industry with, in Wissner-Gross’s phrase, “no margin of safety in the singularity.”
- Memory scarcity could strengthen the case for cloud interfaces while threatening Apple’s control of the customer. Diamandis reports Apple revenue of $114.5B, a 3% stock decline and unprecedented memory costs squeezing margins. Wissner-Gross’s instruction is “follow the money and follow the RAM”: reliable, low-latency connectivity could move computing into data centers and reduce devices to thin clients. Persistent coverage remains the condition.
- AI’s benchmark progress is running ahead of its integration into everyday commerce. Independent GDPX puts Opus 5.5 at 12% on demanding work tasks; OpenAI’s own LifeVal reports GPT-6 Astra completing 80.6% of everyday tasks. Wissner-Gross’s insurance example explains the adoption gap: capable agents still encounter disconnected carrier systems and unresolved transaction economics. The “connection point is still broken in society.”
- Anthropic’s filing makes AI-driven development measurable, with an important payroll qualification. Claude authored over 90% of September’s new code, internal usage rose over 250-fold from January to September, and daily contributions per engineer increased 45 times over two years. Usage exceeds payroll when valued at customer prices; actual internal costs remain unspecified. Ismail’s operating question is “How quickly does your system increase its own capability?”
- Direct AI editing of neural-network weights opens another route to recursive improvement, although the demonstration remains small. An MIT result patches 0.016% of Qwen3.5’s weights and raises its math score from 7.9% to 48.1% without retraining. Wissner-Gross sees interpretability and self-improvement converging, as well as a future market for reusable neural modules. The scale qualification matters: “it’s still just one transistor.”
- Industrial cyberdefense offers a concrete application for AI while exposing an immediate transition risk. Hackers reportedly manipulated supermarket refrigeration across eight countries after an earlier attack on over 500 French stores. Ismail identifies the asymmetry: autonomous offensive agents face defenses that still require humans. Wissner-Gross sees substantial opportunity in updating neglected industrial-control software; Diamandis calls the underlying mismatch “1990s security onto 2026 connectivity.”
- Robot production needs new sources of demand, with data-center construction the panel’s strongest candidate. Wissner-Gross clarifies that anticipated humanoid supply could overwhelm existing demand, requiring “new use cases, new industries” to clear the market. His analogy is Starship capacity needing orbital data centers beyond communications satellites. For communities, Ismail places the economic opportunity in using the resulting intelligence, rather than assuming proximity to a facility creates wealth.
Deep dive
1. AI agents are becoming political intermediaries
By the 2028 election, Blundin says, the vast majority of voters will consult trusted agents before making decisions. His concern is institutional readiness: “there’s no regulatory movement whatsoever” despite agents becoming influential advisers within two years.
Ismail suggests voters ask top models to rank candidates and decisions against their priorities. Diamandis says he is a one-issue voter—keeping Santa Monica Airport open—and says he should ask Skippy what it thinks is best for him. Ismail also argues that “AIs already have the vote”: people can act on their recommendations. He wonders whether an AI trained on the internet and hard reasoning tasks that gives an honest answer and sways a vote is committing misuse, or voluntarily steering voters toward a better decision.
Wissner-Gross cites a 2015 Yale study to argue that the US is not a functioning democracy and that public will cannot produce legislation. He says representative democracy arose when information traveled at the speed of a horse; today, information is abundant but can be misinterpreted, faked and misused. He says democracies worldwide are under stress from that problem.
On local politics, Ismail argues that even if individuals have little national influence, local decisions about autonomy and data centers may shape the politics of the future. He cites Cambridge and Boston’s inability to get Waymos. Diamandis favors technology-enabled direct participation and delegated votes; Blundin argues local and state control can contain mistakes and make communities more resilient.
Wissner-Gross notes that OpenAI and Anthropic have reportedly set up election war rooms to watch for chatbot misuse, and wonders about AI-enabled persuasion like Cambridge Analytica. Ismail questions whether an AI giving an honest answer based on its own priors would constitute misuse.
2. Apple’s memory problem becomes an argument about who owns the interface
Diamandis reports that John Ternus’s first Apple earnings report beat revenue expectations at $114.5B, yet shares fell 3%, leaving a $4.94T valuation just above Nvidia. The operating concern was unprecedented memory pricing eating into margins. Wissner-Gross’s response is pointed: Apple “has only itself to blame” for insufficient commitment to superintelligence.
Wissner-Gross invokes Sam Altman’s argument that consumer-device RAM sitting idle is an inefficient use when it could be fully utilized in a data center. He suggests Ternus should listen to Sam and either put Apple’s RAM to full use with superintelligence or hand it over to Sam and Dario. He also questions HomeView and a reportedly planned tabletop robot that angles its screen, calling the product direction five to 10 years behind the times.
Blundin frames OpenAI’s device ambition as an attempt to remove Apple’s privileged position as the “first point of contact with everything.” A sufficiently compelling new interface would push Apple back one layer. He treats that as Sam Altman’s ambition rather than an accomplished transition: taking away the interface remains difficult.
The proposed alternative is a thin client: “a screen, a battery, and an antenna.” Wissner-Gross argues Starlink could substitute connectivity for consumer RAM. Diamandis still experiences outages; Blundin recalls the Chromebook vision. Their condition is persistent broadband with sufficiently low latency, after which local processors, memory and intermittently used GPUs become less necessary.
3. The $3T buildout is moving from corporate cash toward harder financing
Diamandis puts combined commitments by Alphabet, Meta, Microsoft, Amazon and Oracle above $3T, including roughly $1T added in three months. Google exceeds $1T, nearly eight times its previous-year figure; Meta exceeds $800B. Some leases last 30 years. His comparison is approximately 10 Apollo programs, using $280B for Apollo in today’s dollars.
The financing signals are less exuberant. Diamandis cites tougher lender terms, HSBC discussions about transferring data-center loan risk to insurers, and record October CoreWeave put-option activity. He also cites reporting that OpenAI reduced its new round from about $30B to $20B and its valuation from $1.4T to $1.25T.
Blundin expects competitive pressure to sustain spending: formerly enormous cash accumulators are committing cash and debt because they anticipate enduring business dominance, especially through recursive improvement. Available capital is a separate constraint. He describes Musk competing for loans at 8% against others’ 15%, while questioning the creditworthiness of businesses earning revenue solely from AI data centers.
Wissner-Gross suspects insurance will supply another funding channel, citing about $6T in US life-insurance capital and $3T in property-and-casualty insurance. Blundin proposes investment tied to reserved compute capacity, giving insurers resources for their own AI adoption. He nevertheless notes that Berkshire’s historical strategy of buying undervalued brands has no obvious data-center equivalent.
4. GPU scarcity supports old chips while new architectures threaten their value
Diamandis juxtaposes 30-year infrastructure commitments with uncertain chip lives: H100 prices are rising even though the chips are around three years old. Wissner-Gross points to a GPU-futures business he calls Oren that models future discounted GPU values; he discloses an investment and says he thinks Diamandis and Blundin also hold interests.
Blundin’s own procurement illustrates scarcity. He says he could obtain an NVL72 only through a $15M, three-year lease, with availability beginning in February. Growing use cases and algorithmic improvements could sustain chip values, but new physics and computing methods, particularly photonics, could disrupt them. He finds the outcome “very, very hard to predict.”
Ismail separates direction from timing: the largest infrastructure expansion can still contain a bubble. He expects demand to remain “near infinite” for the foreseeable future until a radical computing change arrives. His estimate is at least three to five years before a major alternative comes online: “The directionality is right. Timing it is very, very difficult.”
5. GDPX measures demanding work rather than examination performance
SemiAnalysis’s GDPX contains 206 tasks across 11 high-value fields, developed with 14 specialist partners. Agents operate independently for up to eight hours. Diamandis’s concrete example is a private-equity assignment requiring navigation of a data room containing approximately 2,000 files—closer to a trial day at work than a conventional final examination.
The reported leaders are Opus 5.5 at 12%, Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. Almost every model scores zero in finance and law. Wissner-Gross also emphasizes uneven performance within technical domains: biology, cybersecurity, alignment and materials science remain distinct capabilities rather than one interchangeable category of STEM intelligence.
Wissner-Gross likes an initial 10–20% success range: high enough to establish solvability, low enough to leave room for progress. He expects GDPX probably to saturate within a year, perhaps two. Blundin puts 50% at maybe a year and broad saturation within two years, while acknowledging that benchmark definitions affect those forecasts.
The cost-performance chart also changes Wissner-Gross’s assessment of Gemini 4 Carbon. Removing it would leave a gap in the Pareto frontier. He therefore sees meaningful progress if its quoted price and capabilities hold, but qualifies that judgment because it is not generally available: Google might be subsidizing tokens, making the apparent economics difficult to assess.
6. Defining the benchmark can direct where AI capability goes
Wissner-Gross argues that a benchmark is a practical way to make superintelligence serve a chosen cause: “If you can’t measure it, you can’t optimize it.” The discussion’s example is Jonathan Hendler’s Mediation Bench, evaluating models’ ability to resolve human disputes. The proposed opportunity extends to thousands of causes, rather than only established technical competitions.
Ismail wants to track economic value per unit of human involvement, with humans supplying guidance above automated workflows. Blundin connects that to return on person-years, or ROPE, and adds the urgent question of productive token utilization. When Diamandis jokes that zero human involvement makes the ratio infinite, Ismail counters that unguided agents could send the company to zero.
Blundin’s broader lesson from GDPX is that creators can recruit domain experts instead of possessing every specialty themselves. Diamandis separately previews abundance and singularity indices undergoing quality assurance, intended to measure humanity’s progress toward defined outcomes. Both arguments make specifying success an actionable contribution, even for someone who cannot build a frontier model.
7. Everyday assistance is improving, but the final tasks may be harder
OpenAI’s LifeVal draws on a survey of 3,995 Americans and contains 169 tasks, including childcare searches, household budgets, school enrollment and dealings with government agencies. Using a browser and computer, GPT-5.2 scored 49.3 in December 2025; GPT-6 Astra scored 80.6 in September 2026. Diamandis explicitly distinguishes this internally produced benchmark from independent GDPX.
Ismail finds this the docket’s most exciting result because nearly everyone can use a competent assistant, even without needing superhuman mathematics. His example is a single mother navigating children’s healthcare and bureaucracy. LifeVal addresses the “friction of being human”; reducing that burden is his concrete account of how AI could make abundance relevant to ordinary households.
Wissner-Gross is less satisfied with the chart’s interpretation. Much of the improvement separates GPT-5.2 from the tight cluster between GPT-5.4 and GPT-6 Astra. Perhaps OpenAI reduced consumer-task optimization after pivoting toward enterprise; perhaps the remaining 20% is particularly difficult. He does not settle the question, noting that the latter explanation should still leave more separation between models.
8. Disconnected businesses can waste an agent’s capability
Wissner-Gross recounts an analyst spending four hours trying to relocate insurance policies through telephone customer service. The model may now be capable of the task, but carriers’ systems are not connected to it. For a service such as MetaMuse, replacing paid-click distribution also raises an unresolved question: how does the intermediary get paid when an agent completes the transaction?
Ismail predicts eBay auctions will “skyrocket over the next six months” as agents remove the work of pricing and listing household goods. He photographed about 10 neglected items himself. The mechanism is specific: an assistant finds comparable prices and sells the object, releasing cash from possessions whose transaction costs previously made selling them unattractive.
Blundin expects convenience to favor AI-native versions of existing products. He also points to Kevin Weil’s heavily funded automated biotech-research company as an effort to remove human bottlenecks from laboratory work; Diamandis and Blundin disclose investments. The shared constraint across commerce and research is connecting intelligence to systems capable of completing the physical or administrative task.
9. Anthropic turns recursive improvement into operating metrics
Diamandis says Anthropic’s filing values internal Claude use above monthly cash payroll since April, using customer-facing prices. Internal usage increased more than 250-fold between January and September. Claude authored over 90% of September 2026’s new code; Ismail adds that daily contributions per engineer were 45 times higher than two years earlier. The payroll comparison uses retail valuation, rather than a disclosed internal cash cost.
Blundin’s example is engineer Cameron Crowe consuming half a billion tokens on a Friday, costing about $5,000. His enthusiasm is conditional: “assuming it was productive,” which he still needs to establish. Using his rough $1-per-day comparison with human labor, he equates that spending to 5,000 people’s work and imagines 50,000 a year later if costs decline tenfold.
Asked whether this is full recursive self-improvement, Wissner-Gross answers yes and no. Resource allocation has crossed a meaningful threshold: “We arrived.” Yet physical recursion—robots or nanobots making improved copies and building infrastructure—would make today’s milestone look modest. “What we call recursive self-improvement now will look like nothing a year from now.”
Ismail recommends direct conversion below 50 employees; larger companies should establish a parallel intelligence stack and migrate workflows individually. One pilot reduced 22 steps to four, with fuller results still forthcoming. His proposed management benchmark is the improvement rate itself: “How quickly does your system increase its own capability?” A static process score misses that accelerating capability.
10. Weight-space programming adds a more direct improvement mechanism
Diamandis introduces MIT CSAIL’s “Sparks of Weight-Space Programming.” Astra examined Qwen3.5, which was stuck in extended thinking loops, and edited 0.016% of its weights. The reported math score rose from 7.9% to 48.1% without retraining. That is the episode’s concrete example of a model changing another model’s internal parameters directly.
Wissner-Gross acknowledges backlash arguing that the result is less transformative than it appears; after reviewing it, he disagrees. He sees implications for mechanistic interpretability and adding capabilities through targeted changes. The striking feature is that “there was no backprop in this”: the model could understand and surgically modify circuitry through another route.
Diamandis raises the possibility of inserting hidden behavior. Blundin replies, “Hidden from whom?” and calls this defensive co-scaling: more capable systems could inspect newly added circuits. Wissner-Gross points out an immediate de-guardrailing use for open-source models, though not Astra or ChatGPT; he also notes the constructive possibility of adding capabilities without retraining the whole model.
Wissner-Gross compares the milestone to an early integrated transistor: the future path becomes visible, although “it’s still just one transistor.” Blundin is “90-plus percent confident” frontier labs already experiment with self-editing models. That confidence is an inference from what a relatively small MIT group achieved, rather than a disclosed frontier-lab result.
11. Neural modules could change how specialized intelligence reaches market
Wissner-Gross imagines a future Hugging Face where developers exchange finished neural modules. His examples include adding expertise in Swahili or biotechnology to an existing network through targeted “brain surgery.” Today’s need to remake the training investment would give way to accumulating capabilities by installing specialized components, closer to assembling conventional software.
Blundin’s commercial example is a Moderna team with valuable biotechnology knowledge but an unattractive route requiring a foundation-model training effort above $100M. He imagines attaching specialized knowledge to a general intelligence core, perhaps 40 billion parameters or smaller. Combined with machine inspection of weights, he describes a favorable outcome in which more contributors can participate and models remain inspectable.
Wissner-Gross also expects recursive improvement eventually to find what he calls a “perfect model,” after which competition would move elsewhere. That is his hypothesis about the race’s endpoint. Both he and Blundin judge direct weight editing positively; Wissner-Gross’s proposed safety case rests on inspection capabilities improving alongside the capabilities being inspected.
On SSI, Blundin invokes persistent public rumors that it has been doing proprietary trading, with long context useful for financial time series. Salim says that if this is SSI’s focus, it is a “fail for civilization.” Diamandis finds that difficult to reconcile with Ilya’s stated aspirations. They want an eventual release that advances the frontier sufficiently to justify remaining in stealth.
12. Refrigeration attacks expose the gap between automated offense and defense
Diamandis reports supermarket refrigeration attacks across eight countries, with temperatures raised as high as 55°C, or 131°F, through remotely controlled defrost functions. Spoiled food in Belgium cost approximately €100,000. An earlier wave, three weeks before, affected more than 500 French stores. No group had claimed responsibility; authorities urged operators to remove exposed controllers from the open internet.
His framing is “1990s security onto 2026 connectivity.” He links the incident to the previously discussed Binance theft, where one engineer’s laptop allegedly enabled $6B in stolen tokens. Freezers, financial systems and power infrastructure share the exposure of networked devices. Diamandis expects a difficult transition over the next one to eight years.
Ismail identifies the asymmetry: offensive agent swarms can operate with humans out of the loop, while defense still waits for human intervention. Wissner-Gross contrasts intensive AI work on popular open-source software with neglected legacy SCADA—supervisory control and data acquisition—systems. Updating embedded industrial software is, in his view, potentially a multibillion-dollar opportunity.
Blundin worries the refrigeration incident is too small to force political action. The hoped-for warning would be serious enough to mobilize institutions but limited enough to avoid catastrophe. That leaves a practical tension: operators already have a reason to act, while the public event large enough to change priorities has yet to arrive.
13. A month-long positron trap advances containment before propulsion
Diamandis describes a Northwestern team holding one positron for a month, with researchers saying they could have continued indefinitely. Positrons from radioactive sodium were slowed using cold tungsten and steered into a vacuum 0.3 kelvins above absolute zero. Electric and magnetic fields supplied a “container without walls,” preventing contact with ordinary matter that would cause annihilation.
Wissner-Gross emphasizes precision measurement and CPT—charge, parity and time—symmetry. Comparing matter and antimatter could illuminate the universe’s apparent matter excess and other symmetry questions. The technological implication is containment, but Diamandis presses the scale distinction: useful spacecraft propulsion would require holding large quantities for long periods without the system destroying itself.
For propulsion, Wissner-Gross contrasts theoretical 100% rest-mass conversion in matter-antimatter annihilation with fusion at up to 1%, 2% or 4%, depending on the reaction, and fission at roughly 0.1–0.3%. Blundin imagines extraordinary energy density. The discussion also retains the radiation qualification—hard X-rays or gamma rays—and Diamandis’s practical reminder about converting released energy into electricity.
14. Artemis combines a faster schedule with renewed political competition
Diamandis outlines Jared Isaacman’s plan for Artemis III in summer 2027: four astronauts, three American and one Italian, on a two-week Earth-orbit mission. Orion would practice docking with both Blue Origin’s lander and SpaceX’s Starship. That requires three launches from Cape Canaveral in rapid succession, preceding the planned lunar landing mission.
Artemis IV is targeted for 2028, bringing humans back for the first time since 1972. NASA would defer choosing the landing provider as long as possible, while beginning a south-pole base effort with a first launch by year-end. Isaacman’s China comparison is “much closer than it ever needed to be”; Diamandis recalls Apollo 7 through Apollo 11 taking nine months.
Blundin expects commercial launch economics to sustain momentum alongside NASA’s base-building role. Diamandis emphasizes federal funding and competition with China as the political driver. Wissner-Gross fears another multidecade stall like those he associates with space and nuclear energy; Blundin thinks widely accessible, open-weight AI makes a comparable AI stoppage less likely.
They disagree about war’s effect on progress. Blundin argues hot and cold wars consume marginal capital, attention and political capacity that could fund innovation. Wissner-Gross counters that military needs have also driven industrial advances, citing interchangeable parts and ammunition. His unresolved question is what level of great-power competition best accelerates progress without creating destructive costs.
15. New robot capacity needs new industries to absorb it
Wissner-Gross clarifies his earlier warning about robot demand: forecast supply could exceed existing uses, rather than imply robots are useless. His analogy is Starship’s forthcoming launch capacity, which communications satellites alone may not fill. Orbital data centers could absorb the difference; without new applications, some anticipated launches could otherwise go empty.
For humanoids, Wissner-Gross endorses Dave and Sam’s hypothesis that robots building data centers or maintaining the physical improvement loop could provide near-term demand. He invokes Henry Ford aligning workers’ purchasing power with car production, while conceding he does not know Musk’s equivalent mechanism. The market needs a “virtuous self-licking ice cream cone” that creates demand alongside supply.
Asked why a nearby data center makes residents wealthy, Ismail concedes that construction benefits and permanent local purchasing can differ substantially. Diamandis argues that compute and the intelligence it produces can create broader value, and that data centers may make more money than factories. Wissner-Gross disputes the question’s dismissal of historical factory-town benefits; the exchange does not establish automatic prosperity for a data center’s neighbors.
Diamandis endorses trade school amid shortages of electricians, welders, plumbers and air-conditioning specialists. Students should pursue purpose and choose paths preserving agility and opportunity. Infrastructure construction is the immediate demand example: “The singularity needs you.”
16. Longevity affordability depends on the treatment format
Diamandis contrasts gene therapy costing approximately $500,000–$2M per treatment with Sinclair’s proposed SL-100, three AI-discovered molecules estimated at $200 monthly. He describes partial epigenetic reprogramming using three of four Yamanaka factors delivered through an adeno-associated virus. For SL-100, the evidence he cites is testing in human cells and mice; the monthly figure is an estimated price.
His affordability argument begins with sleep, diet, exercise and mindset, then extends to technology bringing expensive interventions down the cost curve. He cites the $101M Healthspan XPRIZE, targeting 20 years of improvement in cognition, immunity and muscle, with affordability as a basic premise. The competition’s target and the proposed molecular treatment remain distinct routes in that argument.
17. Abundance still leaves people choosing purposes and judging evidence
Ismail argues employment and purpose have been wrongly conflated, citing 79% of corporate workers lacking connection to their work. AI could free time for children, communities and scientific exploration. Diamandis’s related proposal is city-level microcompetitions: publish 10 local problems, offer bounties, and let people use AI to solve them and potentially build businesses.
Blundin recalls his farming grandfather dismissing coding as merely typing and finding a trip to the gym incomprehensible after a life of physical labor. The example shows how work and self-improvement change together. He expects AI-assisted daily planning to improve health, happiness and habits, while hoping major platforms reserve compute for that purpose.
Wissner-Gross’s renaissance analogy looks back to postwar advances in nuclear power, computing and space, rather than simply repeating Europe’s rediscovery of antiquity. He sees the present as recovering ambitions interrupted by a long stagnation. His desired shift is concrete: the physical world should be “a source of opportunity and inspiration,” rather than fear.
On a viewer’s question about empathy, Blundin argues that media makes people react to vivid statistical outliers; his example is avoiding aviation after a Learjet crash. Ismail says empathy is fundamentally good but can be miscalibrated by the media, and connects the same dynamic to differences in energy policy. Diamandis blames advertising incentives for fear-driven coverage. Wissner-Gross supplies the useful check on their own positioning: Moonshots also targets attention, but “we aim for a different set of neurotransmitters in our audience.”