Escape Velocity — When Does China's AI Stack Break Free?

Escape Velocity — When Does China's AI Stack Break Free? Futurum Research · Long-form analysis The China Question Escape Velocity. When does China's AI stack break free? A reading of the new Futurum 1H 2026 silicon model against Jensen Huang's April interview with Dwarkesh Patel. The numbers and the man with the most to lose are telling the same story. ByErik Bethke, CTO · The Futurum Group CompiledMay 12, 2026 Data sourceFuturum 1H 2026 Refresh (V24) Source Jensen on Dwarkesh Data TFG Semiconductor Practice Jensen Huang is, by any objective measure, the single individual who has the most to lose if China builds a sovereign AI stack. NVIDIA's data center revenue went from $3.3B in Q1 2023 to $48.4B in Q4 2025 — a 14.7x expansion of the most lucrative monopoly position in semiconductor history. The single largest threat to that monopoly is the emergence of viable Chinese silicon at scale. So why is Jensen the loudest voice in tech arguing publicly that the United States is doing this wrong? That's the question that frames everything that follows. In April 2026, sitting across from Dwarkesh Patel, Huang said the quiet part out loud: The policies… resulted in the American telecommunications industry being policied out of basically the world. To concede that market for the United States technology industry is a disservice to our country. Jensen Huang, Dwarkesh Patel interview, April 2026 The framing matters. Huang is not defending NVIDIA. NVIDIA is fine. NVIDIA is the most profitable semiconductor company in history and is on a forecast trajectory to $1.21 trillion of data center semiconductor revenue across the industry by 2030 — with NVIDIA holding the lion's share of merchant GPU. Huang is defending something else: the proposition that denying access to American chips is a finite-duration policy, and the longer it runs, the more it accelerates the very thing it was meant to prevent. The Futurum 1H 2026 Data Center Semiconductors model — published this week off a chip-by-chip, bottom-up reconstruction of every named accelerator shipping in the world — backs his framing with receipts. There is a measurable, dated, financial answer to the question of when China reaches escape velocity. We can decompose it. The Argument What "escape velocity" actually means. Escape velocity, in this essay, is the threshold beyond which further American export controls produce zero capability constraint on Chinese AI. Below the threshold, sanctions impose real costs — slower training runs, fewer FLOPS per dollar, dependence on smuggled H20s or trailing-node Huawei silicon. Above the threshold, China's stack can absorb a complete cutoff without losing a generation of capability. The sanctions still signal. They no longer constrain. Escape velocity is multi-dimensional. The American technology stack runs on at least five vectors, and China has to clear the bar on each independently. We measure them one at a time. "They have datacenters that are sitting completely empty, fully powered." Jensen Huang · on China's energy abundance Vector 1 Energy & grid capacity ETA · ACHIEVED The least-discussed vector. The Futurum model anchors total data center power on a top-down capex envelope: 13.2 GW of deployed AI compute in 2025 scaling to 46.5 GW by 2030, consuming 33 TWh growing to 198 TWh of annual electricity. That's a six-fold increase in AI electricity demand in five years. The constraint, repeatedly, is not silicon — it's substations, transformers, and interconnect queues. The United States is currently waiting five to seven years for new grid interconnects in primary AI corridors. China has the opposite problem: built-out provincial capacity sitting underutilized. Huang told Dwarkesh, in his most quotable line on the topic: "They have datacenters that are sitting completely empty, fully powered." 2025: 13.2 GW 2030: 46.5 GW China grid: built ahead of demand Verdict. Already at escape velocity. China's bottleneck is silicon and packaging, not electrons. The US is the country with a power problem. Vector 2 Talent depth ETA · ACHIEVED Huang's number, on the record: "50% of the world's AI researchers are in China." Whether that figure is precise to the percentage point is debatable; the order of magnitude is not. The Beijing-Shanghai-Hangzhou AI research corridor has matched or exceeded US output on most published metrics — papers accepted to NeurIPS, ICML, ICLR, and (more importantly) system-level contributions to training frameworks like DeepSpeed-compatible kernels, vLLM, and MOE routing infrastructure. The Chinese AI labs — DeepSeek, Qwen team, Moonshot, Z.ai, Baichuan, MiniMax — have published frontier-grade research at a tempo that suggests the pipeline is not pipeline-constrained. It is funded, sized, and self-replicating. The American export-control framework targets compute, not credentials. There is no policy lever that closes the talent gap. ≈50% of global AI researchers Self-replicating pipeline Verdict. Already at escape velocity. This was the vector that was always going to clear first — researchers don't need ASML. Vector 3 Memory (DRAM / HBM / NAND) ETA · 2027–2028 The Futurum chip-share data set is unambiguous on memory: CXMT, the Chinese DRAM champion that most US analysts cannot name, went from $160M of quarterly revenue in Q1 2023 to $3.30B in Q4 2025. That is a 20.6x expansion in eleven quarters. In the same window YMTC (NAND) ran 5.1x, and Nanya (a Taiwanese memory house with significant mainland exposure) ran 4.4x. On HBM specifically — the binding constraint for AI inference — China is one generation behind. SK Hynix shipped HBM3E in volume in 2024; CXMT is shipping equivalents in 2026 with HBM4 on the roadmap for 2027. The gap is real but it is one cycle, not one decade. And the Futurum forecast has all conventional DRAM and NAND resetting in 2028 as HBM4 erodes for HBM5 — a cycle inflection that gives China a natural catch-up window. Our own chip-library reconstruction puts the constraint at the component level. The Huawei Ascend 910C entry — first-party Futurum engineering estimates — captures 8 HBM3 stacks per XPU at 3.2 TB/s aggregate bandwidth and 128 GB per accelerator, on a dual-sourced N6 compute die. Every Ascend that ships consumes 8 HBM3 stacks. Cross-reference that against the Huawei silicon ramp (Vector 4, 13.9x in nine quarters) and the CXMT ramp on the line above (20.6x in eleven quarters), and the math is unambiguous: HBM stack supply, not compute-die supply, is the binding rate-limiter on China's AI buildout. That is also why the CXMT line is the most important line in this paper. CXMT 20.6x · 11Q YMTC 5.1x · 11Q Ascend 910C · 8 HBM3 · 3.2 TB/s HBM gap: one generation Verdict. Escape velocity by 2027 on commodity DRAM/NAND; 2028 on HBM4-class. The cycle reset in 2028 is China's natural catch-up moment. Vector 4 Compute silicon (XPUs & GPUs) ETA · 2027–2028 The hard one. Without EUV access, SMIC is constrained to leading-edge multi-patterning on N7 and early N5. The Futurum chip library captures this constraint at the field-by-field level: the Huawei Ascend 910C compute die is labeled, exactly, SMIC/TSMC — a dual-source strategy that says the quiet part out loud about how much trailing-edge TSMC capacity is still flowing into Chinese designs. And the revenue ramp is undeniable. Huawei XPU silicon revenue went from $340M in Q3 2023 to $4.73B in Q4 2025 — a 13.9x run in nine quarters. That is faster than NVIDIA's GPU ramp from the same starting line, on the same time axis, on trailing-edge silicon. The scaling penalty is real — Ascend 910C delivers worse FLOPS/watt than a Blackwell — but Huang's argument is that China compensates with quantity, not quality: Our chips are better. On balance, our chips are better. There's just no question about it. […] They're not stuck at 7nm, obviously. They're good at manufacturing. They will continue to advance from 7nm and beyond. Jensen Huang on Chinese silicon, April 2026 "Continue to advance from 7nm and beyond." That is the most consequential sentence Jensen said in the interview. He is the chief executive of the company that benefits the most from China being stuck at N7, and he is on the record stating that he expects them not to be stuck. He has access to chip-library data that the Futurum model can only externally reconstruct, and his reconstruction agrees with ours. Huawei 13.9x · 9Q 910C dual-sourced SMIC/TSMC N5 production target: 2027 Verdict. Escape velocity at the threshold of 2028 — when SMIC clears N5 at production yield and Huawei's domestically-sourced silicon clears one-generation parity with NVIDIA's legacy installed base. China will not match Blackwell. They don't need to. Vector 5 Ecosystem & software (CUDA equivalent) ETA · 2028–2030 The slowest vector — and the one Huang is most confident about. CUDA's moat is not the GPU; it is the decade of accumulated framework engineering sitting on top of it. Triton, vLLM, TensorRT-LLM, the CUTLASS template library, the entire ecosystem of optimized kernels for every layer type, every attention pattern, every quantization scheme. Huang put it bluntly: These ecosystems are hard to replace. It costs an enormous amount of time and energy. Jensen Huang on the CUDA moat, April 2026 But here is the inversion: China is the largest open-source software contributor in the world — Huang's words, in the same interview, said as a flat statement of fact. DeepSeek released its models, weights, and training methodology as open source. PyTorch, vLLM, and Hugging Face Transformers are open-source projects that already run on Huawei Ascend via the CANN stack. The question is not whether a CUDA equivalent will emerge in China — it already has, functionally — but whether it reaches feature parity, performance parity, and developer-tooling parity with CUDA itself. This is the lag vector. Software is sticky. Customers don't rewrite kernels because it's hard even when the silicon is better, and Chinese silicon is, on average, worse. The realistic exit is not CUDA-on-Ascend feature parity in 2028. It is CUDA dependency irrelevance by 2030: a critical mass of inference workloads, mostly agentic and mostly running on commodity XPUs, where the cost gap pulls workloads to Ascend faster than the developer-tools gap holds them on CUDA. CANN, Ascend stack functional DeepSeek, Qwen open source Tooling gap closes ~2030 Verdict. Last vector to clear. Escape velocity 2028 on inference-only workloads, 2030 on full training-and-inference parity. The CUDA moat survives the longest — but it survives in a smaller pond. Convergence calculation Each vector measured against the threshold at which further American sanctions impose zero capability constraint. China clears four of five vectors by year-end 2028. Energy & grid cleared 2025 Talent depth cleared 2025 Memory (DRAM/HBM/NAND) 2027–8 2027–28 Compute silicon 2028 2027–28 Ecosystem / CUDA 2030 2028–30 Vector 202520262027202820292030 Cleared The Convergence 2028 is the threshold year. Four of five vectors clear by the end of 2028. Two are already there. Two more cross in the next twenty months. Only the ecosystem vector — the slowest, the stickiest, the one Huang spends the most time defending — extends out to 2030. What does this mean operationally? It means the next twenty months are the last window in which American export controls produce material constraint on Chinese capability. By Q4 2027, SMIC clears N5 production yield. By 2028, CXMT clears HBM4. By 2028, Huawei XPU silicon is at one-generation parity with the NVIDIA installed base — not the leading edge, but the installed base, which is what matters for inference economics. Power is already sovereign. Talent is already sovereign. The American moat closes from five vectors to one. By 2028, sanctions become signaling. By 2030, even the signal is muted. The Counter-Argument What could push escape velocity past 2030. Three things would extend the timeline. First, a genuine choke on ASML — not the current EUV restriction, which already failed, but a multilateral cutoff that includes DUV immersion machines, photoresist precursors, and ion-implantation equipment. The current export control regime targets the top of the stack; a serious one would attack the middle. There is no indication the political coalition exists to enforce this. Second, a step-change in Western silicon that the Chinese stack cannot follow on trailing-edge processes. The most plausible candidate is photonic interconnect — co-packaged optics that drop the cost of scale-out by an order of magnitude and require a fab ecosystem China does not currently have. Real, but probably not before 2029. Third, a regime change in China that voluntarily de-prioritizes the AI build-out. Probability assessment: not in this decade. None of the three are base-case scenarios. They are tail risks to a forecast in which China reaches escape velocity on four of five vectors by 2028 and the fifth by 2030. What Jensen Knows The man with the most to lose has done the math. Return to the question that opened the piece. Why is Jensen the loudest voice arguing publicly against export controls? Because he has run the numbers. He sees the same chip-by-chip ramp curve the Futurum model reconstructs — he has better data; he sells into both ecosystems and sees the orders directly. He has talked to TSMC and ASML in private. He has seen the Huawei Ascend volume from the supply side. He knows, with high confidence, that the present American policy creates short-term revenue protection for NVIDIA in exchange for long-term market loss across the entire US technology stack. And he is willing to say so. "Huawei just had the largest single year in the history of their company," he told Dwarkesh. Not "Huawei could one day catch up." Past tense. Largest single year, already booked, on chips America tried to prevent from existing. The largest argument for the failure of export controls is delivered, on the record, by the chief executive of the company whose dominance those controls were designed to protect. "Huawei just had the largest single year in the history of their company." Jensen Huang · April 2026 The Futurum model puts a number on that "largest single year": $4.73B of XPU silicon revenue in Q4 2025 alone, against $340M nine quarters earlier. The slope of the line is the answer to the question this essay opened with. The Closing Argument Escape velocity arrives in 2028. China does not need to match Blackwell. They do not need CUDA-on-Ascend parity. They do not even need EUV access. They need to clear five distinct thresholds — energy, talent, memory, compute, ecosystem — and they have already cleared the first two and are inside two years of clearing the next two. The fifth, the ecosystem moat, is the slowest, and even that erodes by 2030 against the simple economic argument that XPU inference at 25% of NVIDIA's price pulls workloads off the CUDA stack faster than the tooling gap can hold them on. Escape velocity is not a binary moment. It is the year in which the cumulative cost of denying access to American technology exceeds the cumulative benefit, for both parties. The Futurum data and Jensen's testimony converge on the same answer. That year is 2028. The current export-control window has roughly twenty months to run. After that, the United States is left holding the policy that, in Jensen's words, "policied" American telecommunications out of the world — with the silicon industry next in line. The Chinese stack does not need to be better. It needs to be sufficient. The data says it will be sufficient in 2028. Everything Jensen Huang said in April 2026 was a man telling Washington, on the record, that the clock is running. The data and the man with the most to lose are telling the same story. Conclusion Part II · Pick your hat What are you being asked to believe? A useful investor's lens — what are you being asked to believe? Not what somebody is saying. What hidden premise has to be true for their position to make sense. Strip the assertion, find the belief. The China-AI policy debate looks like one argument. It is ten. Each persona across the data center stack is asking you to accept a different premise — and a different cost. Pick your hat. The action follows. Hat · 01 The NVIDIA Bull Jensen Huang · NVIDIA · sell-side strategy You are being asked to believe Containment is futile. Twenty months of protectionism buys nothing because the Chinese stack reaches capability sufficiency before the window closes. The cost of trying to contain is permanent loss of open-source leadership, ecosystem leadership, and the export revenue that funds the next generation of US silicon. So what Sell the chips. Maximize the duration of NVIDIA's installed-base monopoly while it still exists. Treat sanctions as a tax on US revenue with zero strategic upside. Make every dollar of CUDA dependence stick before the alternative ships at scale. Hat · 02 The AI Safety Hawk Dario Amodei · Anthropic · capability-denial maximalist You are being asked to believe Frontier capability is itself the threat — not the platform, not the ecosystem, the raw compute. Selling chips to China is selling weapons-grade FLOPS to an adversary, and any lag we impose has compounding safety value because frontier AI risks scale super-linearly. The "nukes to North Korea" analogy is not hyperbole. So what Cut off all advanced exports immediately. Eat the commercial loss. Force China to build on N7 instead of Blackwell. Buy time at any commercial cost — every additional generation of training compute denied is irrecoverable for them. The ecosystem lockout is a feature, not a bug. Hat · 03 The National Cybersecurity Hawk CISA / NSA / Cyber Command-aligned You are being asked to believe Every month we extend American CUDA dominance is a month of supply-chain exposure on the largest single attack surface in critical infrastructure. Sustaining the CUDA monopoly into 2030 means subjecting twenty-to-forty-eight months of American enterprise to escalating state-aligned cyber-offensive operations targeting the very ecosystem you're protecting. So what Force the bifurcation now. Accept the cyber tax of a hard split. Better an enforced air-gap that we built than an asymmetric blast-radius we discovered. Own the attack surface deliberately rather than asymmetrically. Hat · 04 The Capability-Denial Analyst DoD / Intelligence Community · strategic patience school You are being asked to believe Even if China reaches escape velocity in 2028, the cumulative capability gap we impose between today and 2028 compounds in our favor across military AI, autonomy, and ISR. The right metric isn't whether the gap eventually closes — it's the integrated lead we accumulate while it remains open. So what Maintain controls regardless of futility. The integral matters more than the endpoint. Every additional month NVIDIA-class compute is denied is a marginal hour of Chinese frontier training delayed. Time is the asset. Compounding lead is the strategy. Hat · 05 The Domestic-Fabs Advocate Intel Foundry · CHIPS Act beneficiaries · industrial policy You are being asked to believe The sanctions regime is the only thing creating sufficient market pull for capital-intensive domestic fabs to ever pay back their CHIPS Act subsidies. If NVIDIA sells freely to China, the demand signal moves offshore and the Arizona/Ohio/New Albany fabs become stranded assets. So what Tighten controls. Use the resulting captive demand to amortize the domestic-fabs investment. Accept higher chip prices as the price of permanent industrial capacity on US soil. Sanctions are an industrial-policy lever, not just a national-security tool. Hat · 06 The Hyperscaler CTO Sundar / Satya / Andy / Mark · custom-silicon strategy You are being asked to believe The China question is a sideshow. The real game is custom silicon — TPU, Trainium, MTIA, Maia — and the policy outcome barely matters because hyperscaler ASICs let us route around both NVIDIA and the sanctions regime. China is a customer problem, not an architecture problem. So what Accelerate the custom-ASIC roadmap. Treat NVIDIA dependence as a transitional state. Stay neutral on China policy because either outcome (open exports or hard cutoff) accelerates demand for your own silicon. The right hedge is to make CUDA optional in your own datacenter. Hat · 07 The TSMC Executive Hsinchu HQ · the irreplaceable substrate You are being asked to believe The geopolitical play is irrelevant to the substrate. TSMC is the only foundry that can ship leading-edge at scale. American sanctions reduce TSMC revenue at the margin — Huawei N6 dual-sourcing notwithstanding — but TSMC's pricing power expands because capacity is finite and demand for both blocs grows. So what Sell to anyone allowed by export rules. Treat sanctions as fungible — every fab block denied to a Chinese customer routes to Apple or NVIDIA at higher ASPs. The substrate wins both ways. Neutrality is the most profitable hat in the room. Hat · 08 The Sovereign AI Buyer UAE · Saudi Arabia · India · France · the middle powers You are being asked to believe Picking a side ends in dependency. The right play is optionality — buy NVIDIA today because it's available, build relationships with Huawei for 2028 when it ships at scale, and use both ecosystems as leverage against each other in the meantime. So what Diversify the stack across both blocs deliberately. Refuse exclusivity demands from Washington and Beijing alike. Build domestic talent and infrastructure that runs on either silicon. The middle powers' Forecast revenue rises from $13.9B to $79B by 2030 (Futurum data). That position is worth fighting for. Hat · 09 The Grid & Energy Investor Vistra / Constellation / Vertiv / utility CEOs You are being asked to believe The chip war is downstream of the energy war. The binding constraint on US AI buildout is not silicon — it is gigawatts. 33 TWh in 2025 grows to 198 TWh in 2030. China's grid is built ahead of demand; America's is not. Whoever owns substations, interconnect queues, and gas turbines owns the build. So what Ignore the chip-policy debate entirely. Buy power infrastructure, transformers, advanced cooling, and grid-connected gas turbines. The bottleneck moves from foundry to fuse box well before 2028. Electrons are the actual moat. Everyone else is fighting on someone else's terrain. Hat · 10 The Wall Street NVDA Holder Long-only equity · the duration-maximizer You are being asked to believe The right horizon is the duration of NVIDIA's monopoly, not the date it ends. Whether escape velocity arrives in 2028 or 2030 matters less than the compounding margin NVIDIA prints between now and then — and the 38% CAGR top-line forecast works in either case. So what Stay long NVDA. Trim around 2027–2028 as Huawei's installed base crosses the parity threshold on inference workloads. Rotate to Broadcom and Marvell as the XPU narrative eats GPU share. The asset isn't NVIDIA — it's compute duration. The choice Ten hats. Ten beliefs. Ten actions — none of which can be done at once. The American policy debate is what it looks like when ten different people, each wearing a different hat, insist the others are wrong. Most of them are not wrong about their own hat. They are wrong about which hat the rest of us should be wearing. Appendix A · Intellectual honesty Other voices on the same question This essay is one voice. It would be intellectually dishonest not to point at the others. Tags: AGREE with our 2028-escape thesis, DISAGREE, SIDEWAYS (different argument, same shelf). Sideways Scoring the Jensen-Dwarkesh debate Noah Smith · Noahpinion · April 2026 Point-by-point debate scorecard. Less escape-velocity, more "who won which exchange." Largest direct audience overlap with this piece. Read Agree Notes on Jensen v Dwarkesh Jordan Schneider · ChinaTalk · April 2026 The China-policy specialist's read. Schneider broadly accepts the diminishing-return framing of export controls; differs on which counter-measures might still work. The most credible parallel voice in this debate. Read Disagree The many contradictions of Jensen Huang Transformer News · April 2026 Reads the same Jensen quotes we use and reaches the opposite conclusion: his contradictions make him an unreliable witness, therefore his China argument is self-serving lobbying. We use the same contradiction as evidence of his credibility (witness against interest). Worth reading the opposing case. Read Disagree China's AI Chip Deficit: Why Huawei Can't Catch Nvidia Council on Foreign Relations · 2025 The substantive bear case. CFR's compute-deficit math implies escape velocity slips well past 2030 and concludes export controls "should remain." Read against our Vector 4 argument — the disagreement is about whether parity is required for escape velocity, not whether the gap closes. Read Agree Two Loops: How China's Open AI Strategy Reinforces Industrial Dominance US-China Economic & Security Review Commission · March 2026 The most analytically rigorous policy paper in the space. Frames China's open-source AI push as industrial-policy infrastructure. Different argument, same destination — and a primary source for the "open-source as moat-erosion" thread we use in Vector 5. Read Agree Silicon Sovereignty: Huawei and SMIC Neutralizing US Export Controls FinancialContent · January 2026 Closest framing to ours. Argues the "Parallel Purchase" policy and SMIC capacity ramp are already neutralizing the sanctions regime. Vaguer on timeline — no 2028 point estimate. Read Agree Stanford HAI 2026 AI Index — "China has erased the US lead in AI" Stanford HAI · April 2026 Authoritative data point on capability convergence at the model layer. Doesn't touch the silicon layer, which is where our argument actually lives — but the model-layer convergence is the leading indicator the silicon convergence follows. Read Disagree The Myth of the AI Race Foreign Affairs · 2026 Argues the framing itself is wrong — neither bloc can achieve "true tech dominance," so the race construct is misleading. We accept the race construct and decompose its end-state by vector. Useful counterweight to remember that the whole framing is contested. Read Sideways Innovation under Pressure: China's Semiconductor Industry at a Crossroads American Affairs Journal · February 2026 Long-form policy treatment. Walks through the SMIC / Huawei / CXMT supply chain in detail without putting a date on convergence. Best read paired with the Futurum data for a full picture. Read Sideways How middle powers can weather US and Chinese AI dominance Chatham House · February 2026 Reads from outside the bloc fight. Argues middle powers (UK, EU, ASEAN) need their own sovereign capacity regardless of which superpower wins. Our Hat #8 is downstream of this analysis. Read Sideways Full Stack: China's Evolving Industrial Policy for AI RAND Corporation · 2026 The mainstream policy-establishment treatment. Less data-driven than our piece, more institutional. Worth reading as the "what Washington's staff is reading" baseline. Read Sources & Notes Futurum 1H 2026 Data Center Semiconductors Refresh (V24) — chip library, vendor market share by quarter, scenario forecasts. Internal: packages/scripts/importPolarisData/csv/aiChipSets_*.csv. PR #3728. Jensen Huang interview, Dwarkesh Patel, April 2026 — direct quotes verbatim. Topics: China energy capacity, 50% of AI researchers in China, Huawei's record year, CUDA ecosystem moat, telecom industry analogy, open-source contribution, SMIC advancement past N7. Huawei Ascend 910C field-level data — compute die manufacturer recorded as "SMIC/TSMC" in the Futurum chip library. Source-of-record: packages/scripts/importPolarisData/csv/aiChipSetschipLibrary.csv and aiChipSetschipLibraryProvenance.csv. CXMT, YMTC quarterly market-share data — Futurum Q1 2023 through Q4 2025 vendor reconstruction. CXMT: $160M → $3.30B (20.6x). YMTC: $510M → $2.62B (5.1x). Energy / capex anchor — Futurum top-down model: 13.2 GW × $35.25B/GW (CY2025) → 46.5 GW × $46.0B/GW (CY2030). Power demand 33 TWh → 198 TWh. Methodology note. "Escape velocity" as defined here is an analytical construct, not an industry term of art. Vector ETAs are point estimates against the Futurum base-case scenario; the bull case shortens each by 6–12 months, the bear case extends each by 12–18 months. The source interview Jensen Huang on Dwarkesh Patel Every quote in this paper is from the April 2026 interview — on China energy capacity, Huawei's record year, the CUDA moat, and SMIC's advancement past N7. Watch the decisive 2.5 hours yourself. Watch · Read transcript The underlying data Get the Futurum semiconductor model Chip-by-chip library, vendor share by quarter, 2030 scenario forecasts — the dataset behind every number in this paper. Available to TFG Semiconductor Practice subscribers. Request access at TFG Futurum Research · Long-form analysis Erik Bethke, CTO · The Futurum Group · May 12, 2026

Subscribe to the Newsletter

Comments