エピソード

  • Intel Foundry Losses, China Export Controls & RTX 5090 Vanishes

    Intel Foundry Losses, China Export Controls & RTX 5090 Vanishes
    2026/10/18
    (00:00:00) Intel Foundry Losses, China Export Controls & RTX 5090 Vanishes
    (00:00:25) Intel 18A Process Progress
    (00:01:24) Intel AMD Stock Rotation
    (00:02:09) China Tech Export Controls
    (00:02:55) Nvidia RTX 5090 Supply Crisis
    (00:03:32) Anthropic Nova Scotia Data Center
    (00:03:58) Key Watchpoints Ahead

    Intel's data centre segment surged 59% year over year to $6.3 billion — its strongest revenue growth in fifteen years. Yet the foundry unit is still deep in the red, and investors rotated out, sending Intel and AMD down 3% this week while Nvidia held steady. That divergence signals a market thesis shift: Nvidia's GPU monopoly is priced as durable, and every other chip play must now prove its economics independently. Intel's new Xeon 6+ server chips, built on the 18A process, launched ahead of internal schedule — a meaningful milestone given the company's decade-long manufacturing struggles. But process execution and foundry competitiveness are different problems. TSMC and Samsung carry a five-to-seven-year capex advantage that a faster-than-expected tape-out doesn't erase.

    On the geopolitical front, China's Ministry of Commerce proposed export restrictions on 40 technologies this week, including 23 AI-related items and chip design tools. A temporary registration requirement is already live; final rules follow a consultation closing after November 15. The US has restricted advanced AI chip exports to China since 2022 — China is now building a reciprocal framework, accelerating the formation of two parallel semiconductor stacks.

    Nvidia's RTX 5090 has effectively disappeared from retail, with secondary market prices reaching $8,500. A Nvidia VP confirmed production continues but supply remains constrained — language consistent with data centre and professional GPU demand absorbing available TSMC capacity ahead of consumer channels.

    Finally, Anthropic signed a 20-year lease for a Nova Scotia data centre campus targeting 420 megawatts by 2028, backed by $2 billion in wind power. Power availability is fast becoming the decisive variable in AI infrastructure geography.

    This episode includes AI-generated content.
    続きを読む 一部表示
    4 分
  • CoreWeave's Debt Crisis, Semiconductor Slide & Taiwan Export Controls

    CoreWeave's Debt Crisis, Semiconductor Slide & Taiwan Export Controls
    2026/10/17
    (00:00:00) CoreWeave's Debt Crisis, Semiconductor Slide & Taiwan Export Controls
    (00:00:42) AI Chip Stocks Rotation
    (00:01:27) Semiconductor Index Eight-Day Drop
    (00:02:02) Taiwan and China Export Control Escalation
    (00:02:50) TSMC Terafab Control Dispute
    (00:03:21) GlobalFoundries Government Stake
    (00:03:54) Key Watchpoints Ahead

    This episode covers one of the most turbulent weeks in AI hardware and semiconductor markets this year. CoreWeave's stock fell twenty-five percent — including a twelve percent single-day drop — as surging bond yields exposed the fragility of the neocloud financing model. With tens of billions in debt and Jensen Huang's endorsement failing to steady the stock, investors are now questioning whether the entities buying AI chips can sustain the debt loads required to keep buying them.

    Intel and AMD both dropped three percent as capital rotated out of AI hardware names, despite Intel posting its strongest revenue growth in fifteen years and AMD's data center division — powered by EPYC and Instinct accelerators — now generating the majority of company revenue.

    The broader semiconductor index fell eleven-point-four percent over eight consecutive down days, partly driven by China's CXMT announcing progress on advanced memory chips — a development that rattled assumptions about Western pricing power in the memory market.

    On the geopolitical front, Taiwan moved to criminalize unauthorized AI chip shipments to China, formally aligning with U.S. export controls. China responded by proposing restrictions on twenty-three AI technologies, including large language models and chip design tools. The global chip trade is fracturing into regional blocs.

    Meanwhile, TSMC's CEO C.C. Wei publicly contradicted Elon Musk's claim that his companies would operate the Terafab facility. And the Trump administration doubled its GlobalFoundries equity stake to eight percent, underscoring a shift from subsidy to direct government ownership across thirty-five semiconductor companies.

    A YesWee production, built using AI technology.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • Broadcom's 221% AI Revenue Surge, XPU vs GPU & TSMC Texas | Ep 1

    Broadcom's 221% AI Revenue Surge, XPU vs GPU & TSMC Texas | Ep 1
    2026/10/16
    (00:00:00) Broadcom's 221% AI Revenue Surge, XPU vs GPU & TSMC Texas | Ep 1
    (00:00:37) Custom XPUs vs Nvidia GPUs
    (00:01:27) Nvidia's Supply Gap Opens the Door
    (00:02:12) The Six-Customer Concentration Risk
    (00:02:54) TSMC Texas and the Terafab Standoff
    (00:03:36) What to Watch Next

    Broadcom's AI semiconductor revenue surged 221% year-over-year to $16.7 billion, with management projecting $115 billion by fiscal 2027 and $230 billion by 2028. This episode unpacks why those targets are credible — and where the risks concentrate.

    At the centre of the story is the accelerating shift from general-purpose Nvidia GPUs to custom XPUs built in partnership with hyperscalers. Anthropic, OpenAI, Meta, and Google are all deploying custom accelerators their teams claim match GPU performance at roughly half the cost. Broadcom designs and supplies silicon for several of these players, giving it a structurally different competitive position than Nvidia's open-market model.

    Nvidia isn't standing still, but it is constrained. The company can fulfil only around 70% of current demand, and memory costs are compressing gross margins to 71–72%. Broadcom trades at 19x forward earnings despite faster AI revenue growth — a valuation gap analysts argue is still pricing the custom silicon story as speculative rather than structural, with price targets implying around 20% upside.

    The concentration risk is real: Broadcom's entire XPU thesis depends on six hyperscaler customers executing on schedule. Only Anthropic is publicly confirmed as the largest 2027 customer. Any power delays or financing pressures could compress guidance fast.

    Separately, TSMC sent evaluation teams to Texas and confirmed talks with Elon Musk's Terafab project. The standoff is stark: TSMC insists it must operate any facility it builds; Musk's model assumes the opposite. With TSMC raising full-year capex guidance to $60–64 billion after reporting record quarterly net profit of $24 billion, its geographic decisions carry enormous consequences for U.S. advanced fab capacity.

    A YesWee production, built using AI technology.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • AMD's CPU Inference Bet, Cerebras Pressure & TSMC Capex Watch

    AMD's CPU Inference Bet, Cerebras Pressure & TSMC Capex Watch
    2026/10/15
    (00:00:00) AMD's CPU Inference Bet, Cerebras Pressure & TSMC Capex Watch
    (00:01:10) CPU vs GPU Inference Battle
    (00:01:55) Capital Constraints Hit Optical Sector
    (00:02:46) Cerebras Valuation Under Pressure
    (00:03:35) TSMC Earnings and Intel Foundry Watch
    (00:04:26) Key Signals to Watch

    AMD's Lisa Su is making a historic call: the CPU total addressable market will exceed $200 billion by 2030, growing at a 50% CAGR driven by agentic AI workloads. Unlike GPU-heavy model training, agentic systems favour CPU-dense architectures — and AMD's Helion rack puts over 32,000 Zen 6 cores alongside 32 MI350 GPUs in a single system. Production deployments at OpenAI, Meta, Microsoft, Oracle, and Anthropic lend the thesis credibility, but Nvidia's CUDA software ecosystem remains the open question that could blunt the entire hardware case.

    Meanwhile, capital constraints are emerging as the new brake on AI infrastructure growth. Broadcom and Intel both flagged funding access — not chip supply — as the binding limit on datacenter expansion. The warning hit optical suppliers hard: Applied Optoelectronics fell 9%, Coherent and Lumentum each dropped 7%. For concentrated suppliers, a capex slowdown isn't just a headwind — it's a structural risk.

    Cerebras posted 281% cloud revenue growth and raised full-year guidance, yet the stock fell 5% as a lock-up expiry and OpenAI's supply diversification spooked investors at a 217x forward earnings multiple. High-growth, high-concentration names have no margin for error.

    On the foundry front, TSMC earnings arrive with focus on whether $60–64B capex guidance executes at the top end and what 2027 visibility looks like. Intel's expanded Applied Materials partnership on 18A and 14A nodes signals foundry ambition — but yield execution remains unproven. CoWoS advanced packaging capacity is the real delivery proxy for AI silicon reaching hyperscaler racks.

    The supply constraint has shifted. Chips are available. Capital allocation is now the variable that determines which AI hardware business models win.

    This episode includes AI-generated content.
    続きを読む 一部表示
    6 分
  • Maia 300 vs Nvidia, TSMC Earnings & the GPU Price Spiral

    Maia 300 vs Nvidia, TSMC Earnings & the GPU Price Spiral
    2026/10/13
    (00:00:00) Maia 300 vs Nvidia, TSMC Earnings & the GPU Price Spiral
    (00:00:58) GPU Consumer Pricing Spiral
    (00:01:48) TSMC October Earnings Stakes
    (00:02:30) AMD Ryzen AI Halo Platform
    (00:03:10) Korea Chip Export Macro Signal

    Microsoft is moving into production with the Maia 300 AI accelerator by 2027, explicitly targeting inference workloads that now represent 71% of AI data center GPU demand. With $37 billion in annual AI revenue on the line, the economics of per-inference cost control are driving one of the most consequential homegrown chip bets in hyperscaler history. Nvidia's training dominance remains intact — but the inference margin is narrowing.

    Meanwhile, the memory squeeze is now fully visible at retail. AMD raised GDDR6 costs for Radeon partners, and Nvidia's RTX 50-series flagship is up 132% year-on-year. The SK Group chair has put hard numbers on the timeline: memory demand will outstrip supply by more than 20% through 2030, with no new fab capacity arriving for four to five years.

    TSMC reports earnings on October 15th, and the stakes are unusually high. HPC chips — AI accelerators and high-end GPUs — drove two-thirds of TSMC's Q2 revenue. Markets are priced for a guidance raise. Any miss on 3nm revenue or capex guidance could trigger a broad sector pullback.

    AMD is running a quieter experiment at the edge with the Ryzen AI Halo Developer Platform — 192GB of RAM, a 55 TOPS NPU, priced at $7,299 — testing whether local agentic AI can sidestep both cloud dependency and the memory bottleneck.

    Finally, the IMF raised South Korea's 2026 growth forecast by 0.7 points to 3.3%, citing semiconductor exports directly. Chip geography is now economic geography.

    This episode includes AI-generated content.
    続きを読む 一部表示
    4 分
  • Lisa Su in Seoul: AMD's HBM4 Crisis, Samsung Foundry Pitch & Nvidia RTX Spark Ships

    Lisa Su in Seoul: AMD's HBM4 Crisis, Samsung Foundry Pitch & Nvidia RTX Spark Ships
    2026/10/12
    (00:00:00) Lisa Su in Seoul: AMD's HBM4 Crisis, Samsung Foundry Pitch & Nvidia RTX Spark Ships
    (00:00:35) AMD-Samsung Seoul Talks Escalate
    (00:01:29) TSMC Concentration Risk in Focus
    (00:02:02) Samsung Vietnam Frees Korea for HBM
    (00:02:46) Nvidia RTX Spark Ships October 16
    (00:03:26) Upscale Token Fabric Watch

    AMD's HBM4 shortage is no longer theoretical — CEO Lisa Su's October visit to Samsung in Seoul signals a full escalation in supply negotiations, with the MI455X accelerator requiring 50% more HBM4 per unit than its predecessor and constraints now expected to run through 2027. This episode breaks down what that means for AMD's production roadmap, its dependence on TSMC, and Samsung's opportunistic pitch to win AMD foundry business alongside the memory deal.

    We walk through the specifics of the AMD-Samsung memorandum of understanding — covering HBM4 for MI455X, DDR5 for Venice EPYC, and the Helios platform — and explain why the October Seoul meeting marks a shift from exploration to concrete capacity negotiation. Foundry switching from TSMC isn't a quick option, but AMD is now having the formal conversation, which itself is a supply-chain signal worth tracking.

    On the Samsung side, a $4 billion investment in chip testing facilities in Vietnam, expected online in Q4 2027, will shift standard DRAM and NAND testing out of Korea's Cheonan plant, freeing it for high-value HBM production. We assess the logic and the unresolved questions around logistics friction.

    Elsewhere: Nvidia's RTX Spark Windows PCs and Surface Laptop Ultra open pre-orders with an October 16 ship date, marking a real milestone for consumer AI compute. Nvidia's unresolved talks with Reflection AI — stake increase, compute provision, or acquisition — remain unsigned and should be treated as optionality. And Upscale AI's Token Fabric, linking AMD, Intel, and ARM chips through Nvidia's Spectrum-X network, begins rollout in Q4 2026, with implications for heterogeneous AI cluster design.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • Nvidia Chip Smuggling Guilty Plea, Korea Capital Flight & d-Matrix

    Nvidia Chip Smuggling Guilty Plea, Korea Capital Flight & d-Matrix
    2026/10/11
    (00:00:00) Nvidia Chip Smuggling Guilty Plea, Korea Capital Flight & d-Matrix
    (00:01:00) Export Controls Actually Enforced
    (00:01:45) Korean Chip Giants Losing Foreign Capital
    (00:02:55) Nvidia's Inference-Layer Consolidation
    (00:03:37) Watchpoints Going Forward

    Export control enforcement just became impossible to ignore. A contractor linked to Super Micro has pleaded guilty to smuggling two and a half billion dollars worth of Nvidia B200, H100, and H200 GPUs to China — routed through server hardware to evade detection. With four criminal counts including export control violations, this case signals a fundamental shift: restrictions on paper now come with real criminal consequences. Co-founder Charles Liaw and a Taiwan-based sales manager remain indicted, leaving Super Micro with unresolved legal exposure.

    The deterrence calculus across the entire AI chip supply chain has changed. Resellers, logistics intermediaries, and foundry partners operating in gray-market channels face a risk environment they can no longer dismiss. A meaningful share of advanced GPU demand from China was flowing through exactly these routes — that window is narrowing fast.

    On the capital markets front, foreign ownership of Samsung Electronics has fallen to 46.38%, the lowest in nearly 19 years — a level not seen since the 2008 financial crisis. SK Hynix tells a parallel story, with foreign holdings dropping below 50% for the first time since 2023. Softening memory pricing, geopolitical risk, and capex uncertainty are pushing institutional investors out of Korean chip equities. Whether this stabilises depends on the next wave of AI infrastructure spending arriving on schedule.

    Finally, Nvidia is quietly consolidating its inference-layer dominance. The company is in early talks to acquire or acqui-hire Reflection AI, while separately disclosing plans to invest in d-Matrix — an inference-chip startup valued at roughly two billion dollars. Pulling d-Matrix into the NVLink ecosystem converts a competitor into a dependent. Classic Nvidia.

    This episode includes AI-generated content.
    続きを読む 一部表示
    4 分