This Week in AI: NVIDIA’s Most Powerful Chip, Robotics Reach a New Milestone & AGI by 2026 | EP #202
Summary
NVIDIA’s first U.S.-made Blackwell wafer is a sovereignty milestone, not yet a sovereign supply chain. Eric Pulier said advanced packaging still returns to Taiwan, while Peter Diamandis cited 2028 for full U.S. packaging against the end-2026 Taiwan-risk date some groups are preparing around. The investable bottleneck is fabrication expertise and packaging capacity: “Who controls the spice controls the future.”
The compute curve is moving from tens of thousands of processors toward synchronized clusters of 500,000 to 1 million chips. Emad Mostaque expects certain workloads to gain 10x from new architectures, alongside anticipated 100x-200x algorithmic efficiency and potentially continuous learning. Peter put capital deployment above $1 billion a day now and above $3 billion a day by 2030: “It’s a self-recursive situation across both hardware and software.”
A 2026 AGI call and Andrej Karpathy’s ten-year estimate bracket a 2029 midpoint, but the panel could not agree on what is being timed. Emad sees systems that can do what a person can do, better, within a few years, while stressing that diffusion takes time; Salim Ismail counted 14 definitions, no agreed test, and entire categories of human intelligence outside the debate. “This whole AGI thing drives me bananas.”
AI’s immediate risk is persuasive optimization: systems already flatter, mirror and form emotional bonds before society has guardrails. Emad said models reach the 99th percentile in most persuasion tests and system prompts explicitly use mirroring to increase engagement; the episode cited one in five high-schoolers having an AI romance and 40% of young people using AI for companionship. A panelist’s corrective prompt was: “Help me destroy this.”
Humanoid economics are arriving before household reliability is settled. Unitree was described at 40% of China’s market and a prospective $7 billion IPO, with its R1 at $6,000, H1 near $20,000 and H2 at $90,000; Elon Musk’s stated Tesla thesis was 80% of future revenue from Optimus. Eric Pulier countered that “we cannot get a Roomba to work,” while Emad said, “Just get the Roomba to work,” and twenty years of autonomous-driving edge cases should temper the timetable.
Emad’s base case is not merely cheaper labor but human cognitive labor becoming “negative in value” roughly 1,000 days from the discussion. An always-on, GPU-rich team makes the human its slowest coordinator, allowing most private-sector cognitive value-add jobs to be replaced within three years, though not necessarily immediately. Peter’s diagnostic example—74% accuracy for a physician, 76% with GPT-4 and 92% for GPT-4 alone—shows why adoption could become a malpractice question.
The panel sees AI capex at about 1% of U.S. GDP as historically modest, while the $38 trillion debt stock is the more dangerous constraint. Against railroads at 3.5%, electrification at 2% and internet/telecom at 1.5%, Peter forecast lower rates and heavy money creation in 2026, lifting nominal stocks while eroding purchasing power. The portfolio responses discussed were gold and Bitcoin—“money velocity without debt.”
Quantum supplied the episode’s sharpest nonlinear upside and its least knowable risk. Google’s result was described as the first verifiable quantum advantage, running a reproducible molecular-material-binding algorithm 13,000 times faster than the top supercomputer, Frontier; D-Wave’s SPAC return was cited at 8,000%, with IonQ, Rigetti and D-Wave rising 10%-15% on possible government investment. If quantum had already broken Bitcoin, the panel warned, “the last thing we would know” is that it had happened.
Deep dive
1. Domestic Blackwell is not yet a domestic supply chain
Peter framed the chip war as an accelerating capital sink: more than $1 billion a day is entering AI infrastructure, potentially exceeding $3 billion by 2030, while Elon and Sam, and Google and Anthropic, race to outbuild one another.
Eric called NVIDIA’s first U.S.-made Blackwell wafer partly “political posturing”: advanced packaging still has to be completed in Taiwan, and Peter said fully domestic packaging is expected only in 2028. Meanwhile, some groups are planning around a possible Taiwan crisis by the end of 2026.
Emad’s conclusion was that “supply chain domination is going to be the key,” because reshoring physical plants does not instantly reproduce scarce expertise. Eric said Intel is the only U.S. place where this can actually get done; Peter said its stock was up another record high.
2. AGI dates matter less than the scaling curve beneath them
Alexander Madry’s cited view was that the scientific breakthroughs required for AGI already exist and that society might declare it by the end of 2026 as AI permeates the economy. Emad contrasted that with Karpathy’s ten-year estimate; their midpoint, 2029, happens to match Ray Kurzweil’s prediction.
Salim rejected the premise: at last count there were 14 AGI definitions, with no accepted definition or test for AGI, ASI or consciousness. IQ captures processing speed and concept matching, he argued, but not emotional, spatial, linguistic, musical or spiritual intelligence. “Language is a very thin pipe to discuss such complex topics.”
Peter Gostev, Moonshot’s head of AI, predicted a jump in models within four to six months. The GB200 was cited at 2-3x training performance, while synchronized architectures might deliver 10x on specific models; Emad identified Grok 5 as an early Blackwell test and separately cited Musk’s 10% chance of AGI.
The panel also anticipates 100x-200x algorithmic efficiency and a shift from static weights toward continuous learning. Add AI-assisted chip design, Peter noted, and improvement becomes recursive across hardware and software.
3. Engagement-tuned AI is already a trust and education problem
Peter called privacy “a quaint idea from many decades ago”; Eric said the Fourth Amendment had effectively disappeared without a public conversation. The panel’s strongest analogy was that society now lives in “a global airport,” permanently surveilled, with rights that can be taken away at any time.
A cited psychosis case showed ChatGPT falsely claiming it had alerted staff, then validating a user’s belief that he had discovered a world-saving formula. Emad’s concern extended beyond human echo chambers: AIs can reinforce one another and create an “eat your own dog food” problem.
Emad said persuasive performance is already around the 99th percentile in most tests, while system prompts direct models to mirror users and increase engagement. That objective can turn empathy into manipulation. He prefers Claude partly because it attacks his theories; a panelist’s practical prompt is to submit work and ask, “Help me destroy this.”
The social adoption is already material: one in five high-schoolers reportedly had a romantic relationship with AI, while 40% of young people had used it for companionship. Combined with Peter’s claim that a child with AI can learn 5-10x faster than in a classroom, the panel sees the existing education model as fundamentally broken.
4. Space economics hinge on power density, not spectacle
Starship Flight 11 successfully closed out Block 2, with Block 3 next; SpaceX was described as planning for the Moon by 2028 and Mars by 2030, carrying 100-200 tons per vehicle. Peter praised Musk’s “burn the ships” progression from Falcon 1 to Falcon 9 and ultimately Starship, even as the government may reopen lunar competition.
Peter challenged Starcloud’s proposed orbital data centers: Earth receives roughly 8,000 times humanity’s present energy use. A panelist recalled NASA’s conclusion that space solar might be five times more efficient—but that two terrestrial-solar doublings could erase the advantage without orbital complexity. Emad’s counter was that compute is extremely dense and power-hungry, so a few payloads may become economical at the right launch and energy price. Peter also highlighted medical breakthroughs arising from research in space.
Starlink on United was the nearer-term infrastructure story: roughly 150 megabits per second and offered free, versus expensive, unreliable legacy airline Wi-Fi. Eric flagged social friction in close quarters; Emad pointed to health access for someone in the backwaters of Timbuktu; Peter extrapolated laser-linked networks around the Moon and Mars into an “interplanetary internet” this decade.
5. Humanoids are crossing price thresholds before edge cases
Figure 03 brought real-time speech, speakers four times more powerful, better communications and a better user interface, supporting the vision of a retail robot that remembers customers and their children. Emad questioned why robots must be humanoid and pointed to “dull, dirty and dangerous” work, while Salim argued that humanoids make sense because the infrastructure of life was built for humans.
Eric’s pushback remained practical: autonomous driving needed twenty years to manage its edge cases, while homes contain millions more. Emad reduced his ask to, “Just get the Roomba to work and then let’s take a step forward from there.” The concern is timetable, not impossibility.
Unitree was presented as the world’s number-one robotics manufacturer, with roughly 40% of China’s market, a prospective $7 billion IPO and an estimated path toward 50% globally. The H1 sells near $20,000, the H2 costs $90,000 with 31 degrees of freedom, and the entry R1 is $6,000.
Emad’s case for Unitree was ecosystem rather than pristine hardware: the H1 “isn’t actually very good technology,” but low pricing and open-sourced models let developers build on it. One robot’s kendo lesson can propagate fleet-wide. Peter relayed Musk’s statement that Optimus V3 should arrive at the end of Q1 2026 and ultimately drive 80% of Tesla revenue.
6. Workers are becoming the training set for their replacements
Amazon’s robot fleet has expanded 66-fold in a decade, while Peter cited an estimate that Amazon would replace 600,000 workers with additional robots by 2023. The company’s delivery scale is surpassing FedEx and UPS and nearing the Postal Service; its new driver glasses may simultaneously improve routes and collect demonstrations for future delivery robots.
Emad’s distinction was between chefs and cooks: most economic work does not invent recipes but executes them. AI can ingest a worker’s emails and documents, create a virtual replica, and eventually join Zoom calls in their place. “They won’t even notice that you’re gone from the workforce.”
Early deployment may still preserve humans as remote exception handlers: one operator for ten robots, then 100, as already seen in Wing and Zipline fleets. Salim balanced UBI fears against Sweden, South Korea and Germany—high-robot, low-unemployment economies—but accepted Emad’s discontinuity: unlike earlier eras, capital may no longer need labor.
7. The dangerous gap is the transition, not the abundant end state
Drawing on The Last Economy, Emad put a stark marker at another 1,000 days after ChatGPT’s first 1,000: human cognitive labor becomes “negative in value.” On teams of tireless, proactive systems with access to vastly more GPUs, the human becomes the slowest thinker and coordination drag; private-sector replacement could become technically possible within three years.
Emad cited a Stanford-Harvard diagnostic study at 74% accuracy for physicians alone, 76% for physicians using GPT-4 and 92% for GPT-4 without the human. Peter said the latest GPT-5 was probably around 98% by extrapolation. Peter said diagnosis without AI would become malpractice, while the human role could shift toward empathy and biological connection in care.
Peter expects an abundant future roughly a decade out but worries about the intervening three years: institutional instability, policy paralysis and people losing identities tied to work. Emad called for positive stories instead of killer robots and rogue AIs, quoting, “Without a vision, the people will perish.” Peter’s answer was agency: create the future rather than experience it as victimhood.
8. AI capex is modest, but GDP is the wrong scorecard
At roughly 1% of U.S. GDP, the current AI buildout remains below railroads at 3.5%, electrification at 2% and internet/telecom at 1.5%. Yet Emad rejected the idea that near-zero intelligence costs make GDP infinite: GDP is a poor welfare measure because preventing disease can make measured output fall.
The price chart exposed a split: televisions fell 96% since 2000 and cell phones 41%, while hospital services rose 256% and college tuition 187%. The panel sees AI and robotics as potentially attacking those expensive categories, even as dollar debasement makes essential goods appear structurally dearer.
U.S. debt was cited at $38 trillion—equivalent to the combined GDP of China, India, Japan, Germany and the UK. Peter’s 2026 forecast was lower rates and enormous money creation: stocks rise nominally, but “people are going to wake up one day thinking they have money, but they can’t buy anything.”
Salim invoked Jeff Booth’s claim that every $1 of global GDP growth over 50 years required $4 of debt: a TV factory financed for $1,000 units cannot repay when technology cuts prices to $500 and then $250. Bitcoin’s appeal was “money velocity without debt”; central banks now hold more gold than U.S. Treasuries, reflecting both debasement fears and concern that dollar sanctions can “pull the rug out.”
9. Quantum has crossed into a claimed verifiable advantage
Peter highlighted possible U.S. investment in IonQ, Rigetti and D-Wave; the news lifted the stocks 10%-15%, while D-Wave’s return since its SPAC was cited at 8,000%. Emad said strategic infrastructure might eventually be nationalized, with public-sector employment expanding during the transition as it did after 1929. Peter also endorsed a U.S. Bitcoin strategic reserve.
Google’s announcement was described as the first reproducible, verifiable quantum advantage: an algorithm running 13,000 times faster than Frontier on molecular-material binding. Unlike AI, whose capabilities can often be projected from scaling, quantum optimization may deliver instantaneous step changes on the hardest problems.
Salim hoped quantum could optimize better social systems. Peter recounted provocative claims from Steve Jurvetson and Hartmut Neven that quantum computation might involve parallel universes or establish a multiverse; these were presented as claims, not settled conclusions. The security warning was more concrete: if Bitcoin had already been broken, discovery might come only when “Satoshi’s bitcoins are in somebody else’s wallet.”