AI Update: NVIDIA’s Record Revenue, Elon’s Data Centers in Space & Gemini 3’s Insane Performance
Summary
Nvidia’s $57 billion quarter makes compute scarcity—not chip demand—the episode’s central investment variable. Revenue grew 62% year on year, 10 gigawatts of AI-infrastructure deals were announced, and Jensen guided to $65 billion next quarter. Blundin’s framing: Nvidia is the “central bank for AI,” minting compute as a currency everybody must buy.
The Nvidia moat remains formidable but no longer looks architecturally singular. Blundin sees another 2–10x performance available as chips are redesigned for AI, supporting repeated 50–70% price increases; yet Google TPUs, AMD, inference ASICs and AI-designed custom kernels could compress development cycles. “We are moving toward a heterogeneous ecosystem,” Wissner-Gross argued.
Microsoft, Nvidia and Anthropic are assembling an AI power bloc whose economics still depend on downstream revenue. Anthropic agreed to spend $30 billion on Microsoft Azure, while Microsoft and Nvidia may invest up to $15 billion; Ismail predicted an eventual Microsoft-Anthropic merger. Wissner-Gross said he would be surprised if Dario allowed Anthropic to be acquired by Microsoft unless Google were uninterested, because Dario and Demis want to remain close and work together. The bubble test is explicit: automation and new scientific markets must justify trillions in capex.
Sovereign AI is emerging as a full industrial stack rather than a procurement category. Saudi Arabia’s stated $100 billion commitment and XAI data-center deal begin with inference compute but could extend through sovereign humanoid robots. Blundin’s qualification matters: Saudi may be a “learning experience,” “test bed” and investment vehicle, while much physical capacity ultimately lands in Texas or space.
Musk’s proposed 100-gigawatt-per-year orbital buildout would move AI’s bottleneck from Earth’s power grid to SpaceX’s launch grid. The plan is roughly one-quarter of average U.S. electricity consumption added every year, contingent on routine Starship operations; Wissner-Gross would be “mildly, if not significantly, shocked” if hundreds of gigawatts were not operating off Earth within ten years. Cooling is presented as solvable through correctly oriented radiators facing the roughly 3-Kelvin cosmic background. The longer arc includes lunar mining and mass drivers, with Diamandis citing a possible 100-terawatt-per-year objective.
Ground infrastructure faces a potentially awkward 2030–2035 power gap before fusion scales. China added roughly 325 terawatt-hours of generation versus about 80 for the U.S.; Google alone plans $40 billion in Texas through 2027, with 6.2 gigawatts of new generation. Diamandis called solar the easiest energy source to scale, while TRISO pebble-bed fuel offers a nearer-term nuclear route that he described as unable to melt down.
Deployment data is becoming the decisive moat across drones, humanoids and AI medicine. Zipline is preparing to build 20,000 aircraft annually after 30 weeks of 15% weekly delivery growth, while Sunday Robotics uses 500 “memory developers” wearing robot-matched gloves to capture dexterity. In medicine, partial epigenetic reprogramming was said to enter human trials in Q1 2026, and Gemini 3 Pro already beats radiology residents on “Radiology’s Last Exam.”
Deep dive
1. Nvidia has turned compute scarcity into monetary power
Diamandis opened with Nvidia’s $57 billion quarter, up 62% year on year, alongside 10 gigawatts of AI-infrastructure deals and an expected $65 billion next quarter. The episode’s macro framing: humanity is at “the beginning of a very long-term buildout” of computing infrastructure.
Blundin traced the company from video-game polygons to neural networks, arguing its chips are only “about halfway down the journey” toward being optimized for AI. He sees another 2–10x architectural gain available, with each improvement supporting price increases of 50%, 60% or 70% because customers will buy the chips regardless.
Blundin sharpened the moat into a monetary metaphor: “Nvidia has just become the central bank for AI,” minting compute as its currency. The exceptional margins may have headroom precisely because every frontier player must acquire that currency to remain competitive.
Wissner-Gross made demand conditional rather than automatic. Trillions in AI capex remain supportable if models automate the existing service economy and create transformative markets through discoveries in mathematics, science, engineering and medicine; if that revenue appears, “the AI compute and NVIDIA capex party can continue indefinitely.”
On the bubble question, Blundin expected some kind of correction because capital is being deployed without regard, but said he did not think the market was a bubble while revenues and valuations were rising in parallel. Wissner-Gross made the test revenue generation: if it continues to scale quickly, it is not a bubble; if it fails, the market looks bubbly.
2. Custom silicon is reopening the architecture contest
Wissner-Gross named the obvious alternatives: Google TPUs powering Gemini, AMD and specialized ASICs for inference-time transformer compute. His destination is not a single CUDA replacement but “a heterogeneous ecosystem of lots of different architectures.”
Blundin said he is not buying Nvidia stock and disclosed an investment in Standard Kernel, which uses AI to redesign kernels and eventually chips for specific algorithms. Google’s advantage is vertical iteration: change the Gemini algorithm, then immediately redesign the next TPU around it—a loop Nvidia does not yet fully own.
Nvidia’s shift from GPUs toward complete AI servers reflects a larger reversal, in Wissner-Gross’s telling. Computing shrank from mainframes to minicomputers, PCs, smartphones and wearables; with Moore’s law ending and “horizontal exponentiation” beginning, the defining computer is again expanding into a coherent data-center-scale supercluster.
3. Anthropic’s alliance turns frontier labs into industrial conglomerates
Anthropic agreed to spend $30 billion on Microsoft Azure capacity powered by Nvidia, while Microsoft and Nvidia may invest up to $15 billion in Anthropic. Blundin contrasted Anthropic’s roughly $300 billion valuation with Microsoft reaching that milestone not long ago: capital at this scale makes previously implausible execution “wide, wide open.”
Ismail predicted an eventual Microsoft-Anthropic merger, reading the partnership as Microsoft diversifying beyond OpenAI and strengthening its enterprise position. Wissner-Gross said he would be surprised if Dario allowed Anthropic to be acquired by Microsoft unless Google were uninterested, because Dario and Demis want to remain close friends and work together.
Blundin also grouped Dario and Demis as leaders whose ethical concerns shape decisions beyond profit maximization, pointing listeners to “Machines of Loving Grace” and Demis’s Nobel acceptance speech. Diamandis retained his belief that Google’s long view—especially under Demis—still includes “what can we do to make the world a better place?”
Wissner-Gross’s structural reading was less about one partnership: Anthropic is becoming another vertically integrated member of the frontier “big N,” with data centers, chip architecture, models and applications. Blundin added that partnerships can deliver acquisition-like control without triggering the antitrust scrutiny that remains big tech’s principal barrier.
4. Saudi AI begins with XAI but points toward a sovereign stack
Diamandis described Saudi Arabia’s $100 billion commitment as part of a long-running ambition to become a global AI superpower—ideally second to the U.S., or third behind the U.S. and China. Making XAI the first customer of an Nvidia-backed facility also supplies the compute Musk needs for Colossus 1 and Colossus 2.
Wissner-Gross sees sovereign inference as only the first layer. As generalist humanoids spread through societal automation, countries may demand sovereign robot compute too: “the rearchitecting from the ground up of an entirely new sovereign stack.”
Blundin’s recent-trip takeaway was more restrained: Saudi’s larger objective may be learning how to invest a trillion dollars intelligently, not becoming the permanent compute capital. He also noted Musk’s parallel paths—early Nvidia supply and a stated $16–45 billion Samsung manufacturing arrangement—while the panel treated a rumored Intel purchase as strategically fascinating but unconfirmed.
5. Orbital data centers would move AI’s bottleneck to launch cadence
In the clip discussed, Musk outlined a path to place 100 gigawatts per year of solar-powered AI satellites in orbit—the lowest-cost route, in his view, to very large-scale AI. The comparison was deliberately jarring: average U.S. electricity consumption is roughly 460 gigawatts, so each year’s deployment would equal about one-quarter of it.
Blundin said Starship would need to become operational on a regular basis, which he expected within 12–18 months. Diamandis described the initial architecture as launching the next generation of Starlink satellites carrying leading chips, unfolding solar arrays, computing in orbit and returning answers through laser links.
Wissner-Gross said the “Overton window just zoomed by” from Dyson swarms as science fiction to an H100 in orbit and prospective multi-hundred-gigawatt clusters. Conservatively, he would be surprised if low hundreds—or many hundreds—of gigawatts of AI compute were not operating away from Earth’s surface ten years from now.
Diamandis also connected the trajectory to lunar mining. O’Neill-style electromagnetic mass drivers could launch lunar-manufactured systems toward Earth orbit; he cited an eventual objective of 100 terawatts of capacity per year, or roughly five times Earth’s total energy output, while describing lunar manufacturing as far out.
The systems-level change is alignment across Musk’s companies: Diamandis praised the ecosystem linking xAI, Tesla and SpaceX, while SpaceX supplies the launch and satellite path for moving the bottleneck “from Earth’s power grid to SpaceX’s launch grid.”
6. Cooling and launch mass—not data links—set the orbital economics
Wissner-Gross rejected the claim that thermally intensive computing requires conductive cooling. Space’s cosmic background is roughly 3 Kelvin; radiators must simply face cold sky rather than the Sun. Orientation is a navigation problem, he argued, but radiative cooling is “completely doable.”
Blundin cited the first H100 placed in space two weeks earlier: about 50 kilograms, cooled with aluminum rather than exotic metals. Most of that mass was sensors and supporting equipment, making weight reduction the nearer-term investment problem before any lunar manufacturing system becomes relevant. He also cited a potential sixfold increase in energy density in orbit after accounting for day-night cycles and solar flux.
The launch-cost curve is already dramatic: the panel compared roughly $1 billion for a shuttle mission with $50–100 million for SpaceX and a $5–7 million target at Relativity Space. Diamandis relayed Musk’s estimate that methane sourced beside a launch site, plus solar-produced oxygen, could eventually make an orbital seat cheaper than a transatlantic ticket.
7. Earth’s power race has a dangerous gap before fusion arrives
Diamandis’s chart showed China adding roughly 325 terawatt-hours of electricity generation over the referenced year versus about 80 for the U.S.—approximately a fourfold difference. Ismail focused on the “big red drop in coal” alongside China’s leadership in solar, wind, nuclear and gas.
Diamandis’s economic claim was categorical: solar capex plus opex is now cheaper than the opex alone of fossil generation. That should drive fossil facilities out of ordinary use, though dense-energy applications may persist. Blundin called for a U.S. “Operation Warp Speed for energy.”
Diamandis challenged the comforting long-term view: first fusion plants may arrive around 2030–2035, but mass production may not come until 2040. Blundin expects supply to remain manageable through 2030 because chip production itself constrains demand, followed by a “massive gap” through 2035 as fabs accelerate ahead of generation.
Google’s $40 billion Texas plan through 2027—including 6.2 gigawatts of new generation and a $30 million energy-impact fund—shows the terrestrial response. The panel credited Texas’s land, launch access and friendly regulatory environment; Blundin argued its lead should pressure Ohio, Wyoming and other states to compete.
8. TRISO fuel brings a decades-old nuclear design into production
X-Energy’s construction of a Category II nuclear-fuel facility matters because it manufactures TRISO fuel for the next generation of small modular reactors. Wissner-Gross’s headline was simple: “Pebble-bed nuclear is finally happening” after decades of delay.
His analogy was a gumball machine filled with billiard-ball-sized spheres containing uranium particles inside carbon-ceramic material. Helium passes through the bed, carries heat to electricity generation, and the pebbles decay over roughly three years before removal and recycling.
Diamandis emphasized the safety claim: these reactors “literally cannot melt down” in the manner associated with Fukushima or Three Mile Island. Blundin’s broader framing connected the revival to AI as humanity’s “innermost loop,” pulling energy, space, manufacturing and previously shelved scientific concepts forward together.
9. Zipline turns drone delivery into a compounding data network
Zipline is expanding South San Francisco production toward 20,000 autonomous aircraft annually. CEO Keller Clifton said deliveries had grown about 15% week over week for 30 consecutive weeks, with an autonomous delivery every 30 seconds and a new Walmart Supercenter launched weekly across Dallas.
Diamandis preserved the company’s regulatory-arbitrage origin: it began in Rwanda and Ghana delivering blood and critical medical supplies, improved through early failures, and operated those networks from South San Francisco. It now performs roughly one million commercial deliveries annually after flying 70–100 million autonomous miles.
Wissner-Gross expects mature drone logistics to relocalize supply chains and support hyperlocal manufacturing—the first practical version of the flying-car future. Blundin added that drone price-performance had at one stage doubled every nine months, while China was already delivering coffee and other goods through the air.
The secondary asset is planetary observation. Autonomous cars, drones, AR glasses and satellites could make “everything knowable on the planet” continuously searchable; the panel cited roughly one gigabyte per second captured by each Waymo car. Privacy is the cost, while detecting illegal fishing, deterring poachers and locating disaster victims are the countervailing benefits.
10. Humanoid competition is becoming a race for training data
Diamandis showed Unitree’s roughly $16,000 G1 performing household chores and noted the company was about to go public, while preserving Musk’s claim that Optimus Gen 3 could become “the single most useful machine on the planet.” The expanding field spans conventional motors, hydraulic tendon systems and non-humanoid form factors.
Sunday Robotics has recruited 500 “memory developers” who wear gloves matched to its robot hands while doing ordinary work. The resulting demonstrations—loading delicate glasses and folding socks—are claimed to form the largest robotic dexterity dataset, without teleoperated training data.
Wissner-Gross said vision-language-action models provide the algorithmic framework; the unresolved problem is data. Robot-matched gloves, Figure’s palm cameras and sim-to-real synthetic training are competing approaches, and he expects one or more to “solve robotics imminently.”
Blundin stressed the deployment flywheel: a robot placed into a narrow use case generates new data every day, retraining the model nightly and widening its lead. Ismail remained more cautious about general-purpose home work such as handling salad plates, expecting dangerous and industrial jobs to mature first.
11. Electric-vehicle forecasts still mistake exponential curves for lines
Ismail attacked the International Energy Agency’s projection of almost no Indian EV adoption through 2035 as “complete horseshit,” expecting the curve to turn vertical. He recalled an earlier forecast that a million EVs would take until 2040, only for Tesla to have one million vehicles on the roads by the time the report was finished.
His mechanism was mechanical simplicity rather than policy: roughly 2,000 moving drivetrain parts in a combustion vehicle versus 17 in a Tesla. Design, reliability and maintenance economics become increasingly difficult to resist, with autonomous electric cybercabs adding another displacement force.
12. AI medicine is moving from assistance toward biological intervention
Diamandis said Life Biosciences plans human trials in Q1 2026 for partial epigenetic reprogramming using a modified adeno-associated virus to deliver OSK genes. After demonstrated age reversal in mice and monkeys, the first targets are NAION—described as “strokes in the eye”—and glaucoma, before possible expansion to liver and other organs.
The reasoning is that youth lacked the diseases later associated with an altered epigenome; restoring an earlier state might therefore remove them. Diamandis kept the implication conditional: if this is genuine age reversal, its relevance extends beyond any single organ.
Wissner-Gross cited Dario Amodei’s public expectation that disease biology and medicine could be solved by the end of the decade. Anthropic’s hiring of life-science researchers and construction of dedicated capabilities is, in his view, what a serious five-year attempt to solve biology looks like.
On “Radiology’s Last Exam,” roughly 50 images across modalities and body systems, Gemini 3 Pro beat radiology residents but not yet board-certified specialists. Extrapolating the line from GPT5 thinking to Gemini 3 Pro, Wissner-Gross predicted that “radiology” would be solved in the next year; Blundin’s radiologist brother-in-law welcomed that prospect because the objective is saving lives, not preserving tasks.
13. Abundance becomes tangible when AI remediates physical scarcity
Wissner-Gross treated Rainmaker’s drone-based plan to refill the Great Salt Lake as an early preview of biosphere-scale remediation. The immediate problem is not merely falling water levels but arsenic from the exposed salt bed becoming aerosolized; automation could make such labor-, capital- and energy-intensive interventions feasible.
Diamandis reframed water scarcity around inventory: 97.5% of Earth’s water is saltwater, 2% is locked in ice and humanity fights over the remaining half-percent, while quadrillions of liters sit in the atmosphere. Ismail defended climate engineering bluntly: industrial society already geoengineered the atmosphere, and “nation-states cannot solve climate change” without technology.
A cited $1.4 billion ReElement/U.S. government initiative aims to rebuild a domestic rare-earth supply chain. Wissner-Gross extrapolated toward hyperlocal nanotechnology scavenging nearby feedstocks within 10–15 years; Blundin already has investments using AI vision and sensors to recover valuable materials from recycling and trash.
Ismail closed with an “epic steamroll towards the era of abundance”: collapsing costs, industries converging and new markets forming between them. His consumer endpoint is personal AI functioning as doctor, lawyer, tutor, mentor and coach—“and it’s all going to be free.”