
Hosted by YesOui · EN

(00:00:00) TSMC Bottleneck Breaks Open: Intel, Nvidia & HBM4 Reshape the AI Supply Chain (00:00:42) Nvidia Tests Intel 18A Process (00:01:29) TSMC CoWoS Sold Out, 2nm Ramp (00:02:04) SK Hynix HBM4 and Nvidia Vertical Integration (00:02:53) AMD UK Sovereign AI Push (00:03:22) Broadcom $35B Infrastructure Platform (00:03:53) Key Watchpoints Ahead TSMC's CoWoS advanced packaging capacity is now sold out with no near-term relief, and that single constraint is reshaping the entire AI hardware supply chain. In today's briefing, we unpack what that bottleneck means for hyperscalers, chip designers, and investors tracking the infrastructure layer of AI.Google has committed to over three million TPU chips from Intel's foundry through 2028 — roughly half of Google's projected TPU output for that year. It is the clearest validation yet that Intel Foundry has moved from aspiration to execution. At the same time, Nvidia is running multi-project wafer tests on Intel's 18A process node for a four-die GPU design tied to the next-generation Feynman architecture. Evaluation is not commitment, but hardware validation at this scale is a meaningful signal.TSMC's Q1 results confirmed the pressure driving all of this: $35.9B in revenue, 66.2% gross margins, and net income up 58%. The 2nm volume ramp in 2027 is now the critical path for next-generation AI GPUs.SK Hynix is supplying HBM4 memory across all four of Nvidia's next-gen platforms using MR-MUF bonding targeting 16-layer stacks in H2 2026 — an ambitious and unproven spec at volume. The SK-Nvidia relationship has moved well beyond transactional supply toward something closer to vertical integration.AMD announced a multi-billion pound UK initiative spanning Dell, Cambridge, Imperial College, and the UK Atomic Energy Authority. Broadcom, Apollo, and Blackstone launched a $35B platform targeting 20-plus gigawatts of global AI infrastructure, with the first tranche funding Anthropic's capacity expansion via Fluidstack.Watch Intel's 18A yield data and TSMC's 2nm ramp in H2 2026. Those are the two metrics that will determine how much of this diversification story actually delivers.This episode includes AI-generated content.

(00:00:00) Google & Nvidia Trial Intel 18A, HBM4 Race Heats Up & Export Control Friction (00:01:30) Intel 18A Yield Risk (00:01:55) TSMC Market Cap Surges (00:02:26) HBM4 Memory Supply Race (00:03:24) AMD UK and Export Control Friction (00:04:17) Key Watchpoints Ahead Intel's stock surged twelve percent this week after Google placed orders for over three million TPU chips using Intel packaging and Nvidia launched early 18A process trials for multi-chip designs. Two of the most important names in AI hardware are now putting real design work through Intel's foundry — a meaningful shift from where Intel Foundry stood just eighteen months ago. The catch: trials are not production ramps, and Intel's yield execution at this node is still unproven.Meanwhile, TSMC's market cap crossed one point four two seven trillion dollars — up over one hundred percent year on year — cementing its position as the backbone of every major AI training cluster on the planet. That concentration is precisely why Google and Nvidia are stress-testing Intel as an alternative.In memory, the HBM4 race accelerated sharply. Samsung disclosed its HBM4E architecture reaching sixteen gigabits per second per-pin and four terabytes per second of bandwidth, while entering foundry talks with Nvidia at four and eight nanometer nodes. SK Hynix countered with a major TC bonder equipment order for HBM4 production at its new M15X fab, with its CEO confirming SK Hynix as Nvidia's largest memory supplier across Vera Rubin, RTX Spark, and Jetson Thor.On the geopolitical front, AMD committed two billion pounds to UK AI infrastructure over five years. And Nvidia CEO Jensen Huang declined Senator Elizabeth Warren's invitation to testify on China export controls — a posture that signals Nvidia is betting on commercial access rather than regulatory compliance theater. The export control policy trajectory remains genuinely unclear, and its resolution will reshape every foundry and memory supply chain move in this space.This episode includes AI-generated content.

(00:00:00) CXMT's HBM3 Breakthrough, Memory Bottlenecks & Broadcom's Punished Growth (00:00:32) CXMT Scale Ambitions and Real Limits (00:01:25) SK Hynix's Strategic Position (00:02:02) Memory as the New AI Bottleneck (00:02:46) Broadcom Guidance and Market Discipline (00:03:27) What to Watch Next China's ChangXin Memory Technologies has confirmed HBM3 manufacturing capability, compressing what was a three-year technology gap with South Korea into something far more strategically significant. CXMT is targeting 60,000 HBM3 wafers per month by end of 2026 and planning a $4 billion IPO to fund the buildout — but the critical question is whether yield data will confirm commercial viability, not just technical achievement.Meanwhile, Jensen Huang's visit to Seoul wasn't incidental. Nvidia's Vera Rubin platform enters full production in Q3 2026 and requires HBM4 from all three qualified suppliers — SK Hynix, Samsung, and Micron — with SK Hynix holding the dominant allocation. Nvidia is locking in supply before demand peaks, embedding Korea's structural advantage into its next product cycle.The broader shift: AI's scarcity bottleneck is moving from GPU availability to memory bandwidth. HBM4 is the rate-limiting factor for frontier model training and inference at scale — which reframes CXMT's ambitions as a geopolitical lever, not just a commercial one.On the market side, Broadcom reported $10.8 billion in AI revenue with 180% year-on-year growth — and its stock fell 7.5%. Investors punished the company for not raising its custom AI chip target above $100 billion for fiscal 2027. The AI hardware cycle is maturing into a market that demands guidance precision over momentum.Key near-term signals to watch: CXMT yield data, Nvidia Vera Rubin customer commitments, and Jensen Huang's Senate testimony on June 11 regarding military diversion allegations tied to China exports.This episode includes AI-generated content.

(00:00:00) ARM Enters Chipmaking, TSMC Locked Through 2027 & Google's $30B GPU Deal (00:00:58) TSMC Shortage Through 2027 (00:01:56) Vera Rubin HBM4 Supply Locked (00:02:46) Google's $30B GPU Signal (00:03:16) Taiwan's Manufacturing Leverage (00:03:46) ARM's Unresolved Question ARM has crossed a line it cannot uncross. In March 2026, the company announced its AGI CPU — its first move into actual chip manufacturing, not just licensing. Orders have doubled, but production is constrained, and the competitive implications for Qualcomm, Apple, and Nvidia are significant: ARM now competes with the very companies it licenses to. The incentive structure shifts in ways that are hard to walk back, and ARM's stock — up roughly 84% year to date — prices in a clean ramp that is far from guaranteed.Underpinning all of this is a TSMC-confirmed AI chip shortage running through 2027, with advanced-node pricing set to rise 3–15% in 2026. TSMC is managing scarcity deliberately, using it as a revenue lever rather than reacting in panic. Meanwhile, two-nanometer production remains anchored in Taiwan through the end of the decade — the $250 billion US investment deal notwithstanding. Taiwan retains roughly 90% of advanced-node output, a concentration that is strategic, not accidental.On memory, Jensen Huang confirmed HBM4 supply is secured across Samsung, SK Hynix, and Micron for Q3 2026 shipments supporting Vera Rubin. SK Hynix holds 60–70% of allocation. The longer-term capacity expansion at SK Hynix's Yongin fab doesn't begin equipment installation until February 2027, making the near-term supply gap real even as the roadmap looks credible.The most striking data point: Alphabet agreed to pay SpaceX $920 million per month for 110,000 H200 GPUs through June 2029 — a $30 billion commitment that signals Nvidia's direct allocations are not meeting hyperscaler demand. Cloud leaders are sourcing GPU capacity through routes that weren't in the standard procurement playbook.This episode includes AI-generated content.

(00:00:00) TSMC-Nvidia AI Fabs, HBM Thermal Wars & Warren's China Probe (00:01:11) FabTwin Digital Factory Simulation (00:01:39) HBM Thermal Wars Samsung SK Hynix Micron (00:03:07) Warren Targets NVIDIA China Testimony (00:03:44) TSMC Geopolitical Risk Repriced The AI hardware and semiconductor industry hit several major inflection points in today's briefing. TSMC and Nvidia formalized their AI-in-manufacturing partnership at GTC Taipei, deploying cuLitho for 20–50% lithography cycle-time improvements, cuEST for material simulation running 50x faster than conventional methods, and the Metropolis platform for computer vision inspection across the fab floor. Nvidia is no longer just TSMC's largest customer — it is now a tooling partner embedded inside the foundry's operations. TSMC's FabTwin digital factory, built on Nvidia Omniverse, extends this further by simulating tool layouts and production flows before a single dollar of physical capital is committed.In memory, the HBM5 thermal race is now three-way. Samsung's Heat Path Block technology cuts thermal resistance by 16% at the die-to-die interface. SK Hynix claims 30% thermal resistance reduction with integrated cooling elements embedded in the HBM package. Micron is pursuing through-silicon via trench cooling — a patented passive vertical heat-management structure. With next-gen AI server GPUs approaching 1,000 watts per chip and HBM stacks trending toward 20 layers, thermal management is no longer peripheral. It is the constraint.On the regulatory front, Senator Elizabeth Warren formally requested Jensen Huang testify before the Senate Finance Committee on Nvidia's China strategy and export-control compliance, with an active House probe running in parallel. Tighter controls could materially impact Nvidia's China revenue.Finally, TSMC shares fell 7% in a single session following the Trump-Xi summit — a sharp reminder that the foundry's 70% share of advanced global capacity carries concentrated geopolitical tail risk that markets periodically reprice.This episode includes AI-generated content.

(00:00:00) RTX Spark vs x86, TSMC Price Hikes & PCB Supply Crisis | Computex 2025 (00:00:46) Intel's ARM Compatibility Warning (00:01:19) TSMC Pricing Power Formalized (00:02:11) Intel Crescent Island GPU Revealed (00:02:48) AMD AM5 Extended; PCB Crisis Deepens (00:03:44) Senate Hearing and Watchpoints Nvidia made its most aggressive move into client computing in years at Computex, unveiling the RTX Spark Superchip — a Grace CPU and Blackwell GPU combination targeting Windows ARM laptops and desktops. Microsoft, ASUS, HP, and Lenovo are all on board for a fall launch. Intel responded publicly, flagging DRM compatibility risks and software fragmentation as its defensive line against Nvidia's x86 challenge.But the structural story of the week may be TSMC. CEO C.C. Wei confirmed 15% price increases on N3 nodes arriving by late 2026, with N2 capacity sold out through 2027. Overseas fab costs in Arizona, Japan, and Germany are the driver. Every major AI chip — Nvidia Blackwell, AMD's latest — runs through these nodes, meaning the cost pressure lands in chip pricing, margins, or timelines. TSMC is projecting hyperscalers will spend $725 billion on AI infrastructure in 2025 alone.Intel disclosed further details on Crescent Island, its Xe3P data center GPU featuring 480GB of memory, positioned as a memory-bandwidth answer to Blackwell for agentic AI workloads. AMD extended its AM5 platform commitment through 2029 and revived the 5800X3D line with a re-engineered silicon bonding process.On the policy front, Senator Elizabeth Warren has summoned Jensen Huang to testify before the Senate on export controls and Nvidia's China exposure. And the U.S. PCB supply crisis deepened — domestic share has fallen from 30% to 4% of global production, with prices up 40% in just two months. TTM Technologies and Sanmina are expanding U.S. capacity, but the gap is enormous.A YesWee production, built using AI technology.This episode includes AI-generated content.

(00:00:00) DRAM Pricing Shock, Grace Blackwell Ships & Edge AI Heats Up (00:01:06) CXMT and the Medium-Term Risk (00:02:01) Grace Blackwell Hits Commercial Stage (00:02:35) Snapdragon and AMD at the Edge (00:03:29) ASML's Equipment Position South Korea flipped its trade balance with China to a $3.8 billion surplus in a single month, powered almost entirely by memory chips. First-quarter 2026 DRAM contract prices rose 93–98% quarter-over-quarter — a pricing shock, not a pricing uptick — as AI data centre demand drains global supply. SK Hynix and Samsung are capturing demand that once flowed to Japan and Taiwan, and their near-monopoly on high-bandwidth memory positions them for further gains into 2027.But the window has a closing date. TrendForce identifies Chinese DRAM producer CXMT as the structural threat. Capex disclosures and yield data remain opaque, making the ramp timeline genuinely uncertain. Investors modelling sustained pricing power through 2027 should treat that as a base case with real tail risk — particularly in HBM, where qualification cycles, not just volume, are the moat.At Computex 2026, NVIDIA's Grace Blackwell architecture moved from roadmap to retail. ASUS launched the Ascent GX10, a petaflop-class enterprise edge system on the GB10 chip with 128 GB of unified memory. The same show saw ASUS debut its first Qualcomm Snapdragon X2 Elite AI Mini PC and an AMD Ryzen AI all-in-one desktop — three architectures, one OEM, one week. The edge AI platform race is now a live commercial competition.Rounding out the picture: JPMorgan reiterated its buy on ASML after the company posted $10.15 billion in Q1 revenue with a 48.7% return on equity, confirming that equipment makers remain the structural beneficiaries of AI scaling regardless of which chip architecture wins.This episode includes AI-generated content.

(00:00:00) Intel 18A Yields, Micron's $1T Moment & Nvidia's $91B Warning (00:00:49) Rackscale Agentic Infrastructure Play (00:01:41) Micron's $1 Trillion HBM Signal (00:02:30) Nvidia Supply Warning and $91B Guidance (00:03:03) US Closes China Chip Export Loophole (00:03:33) ASML and Korean Wage Precedent Intel's 18A process node moves from roadmap to revenue this week, as the Xeon 6+ begins shipping with over 36,000 cores per liquid-cooled rack — a direct challenge to GPU-centric data centre orthodoxy at exactly the moment agentic inference workloads are shifting CPU-to-GPU ratios back toward parity. A rackscale partnership with SambaNova and Foxconn puts disaggregated inference into commercial deployment, with Together.ai already running live customer workloads across Intel, SambaNova, and Nvidia Blackwell hardware.Micron crossed $1.095 trillion in market cap this week, its Cloud Memory unit posting $5.28 billion in revenue at a 66% gross margin — accelerator-grade economics applied to memory, driven by insatiable HBM demand across every major AI accelerator platform. The margin risk if supply constraints ease by 2027 is real and worth modelling now.Nvidia guided approximately $91 billion in Q2 FY2027 revenue while simultaneously warning that bottlenecks in memory, optics, and final assembly may persist through fiscal 2027 — a combination that reflects both genuine demand excess and deliberate demand management.Also covered: the Commerce Department formally closing the subsidiary loophole allowing Chinese firms to source advanced AI chips through offshore affiliates; ASML's $10.3 billion Q1 revenue with a 45-billion-euro EUV backlog signalling a long capex runway; and a Samsung–SK Hynix wage deal setting a new labour cost floor across Korean chipmaking with structural implications for global memory margins.The watchpoints: Intel's hyperscaler qualification timeline on 18A, HBM pricing stability through 2027, and whether US export enforcement closes real gaps or simply redirects them.This episode includes AI-generated content.

(00:00:00) BIS Closes Subsidiary Loophole, Dell +88% & Arm Cloud Surge (00:01:30) Enforcement Gap Still Incomplete (00:02:13) China's Domestic Chip Response (00:02:56) Dell's AI Hardware Surge (00:03:18) Infrastructure Certification and ARM Gains (00:03:57) What To Watch Next The Commerce Department's Bureau of Industry and Security issued new guidance on June 1st, closing a critical loophole that allowed Chinese-headquartered companies to acquire advanced Nvidia Blackwell chips through overseas subsidiaries operating outside mainland China. The fix arrives after a 13-month enforcement gap created when the Trump administration suspended the Biden-era AI Diffusion Rule — a window during which industry estimates suggest hundreds of thousands of top-tier AI accelerators may have reached Chinese-controlled infrastructure.Today's episode breaks down exactly what the new BIS guidance does — and what it doesn't. The subsidiary loophole is addressed, but enhanced due diligence requirements for foundries such as TSMC, designed to catch front-company purchases, were not restored. That leaves a separate diversion route open, and retroactive liability for purchases made during the suspension window remains unresolved.On the supply side, China's domestic chip ecosystem is responding. Huawei Ascend, Alibaba, Moore Threads, and Cambricon are all accelerating development of domestic AI accelerators as the strategic cost of foreign dependency rises.Beyond the geopolitical story, Dell reported an 88% quarterly revenue jump driven by AI server demand, with stock rising 33%. Nutanix Unified Storage received Nvidia GPU-direct certification, reducing a key storage bottleneck in enterprise AI deployments. And Arm-based cloud processors — AWS Graviton, Google Axion, Azure Cobalt, and Oracle Ampere — are delivering roughly 60% energy efficiency gains, with Spotify, Pinterest, and Uber already shifting workloads.Key watchpoints: whether BIS restores foundry screening requirements and whether retroactive enforcement follows.This episode includes AI-generated content.

(00:00:00) US Closes China Chip Loophole, AMD Venice 2nm Ramps & EU Fab Push (00:00:59) US Closes China Chip Loophole (00:02:04) EU Sovereign Chip Fab Push (00:02:42) Qualcomm Edge AI Expansion (00:03:25) Data Centers as Grid Assets The US Commerce Department has moved to close a significant regulatory gap that allowed Chinese-headquartered entities operating outside China to receive advanced semiconductors — including Nvidia and AMD silicon — without triggering licence requirements. Estimates suggest hundreds of thousands of chips may have moved through that window over roughly a year. The new guidance is real progress, but a structural gap remains: foundries like TSMC still face no obligation to verify their customers' customers, leaving secondary procurement vectors open.On the silicon front, AMD's Venice EPYC processor is now in production on TSMC's 2nm node — no longer a roadmap promise, but an active ramp. With Nvidia Blackwell and AMD Venice both converging on TSMC 2nm, the foundry's grip on leading-edge AI-capable silicon is tightening. AMD's next step, the sixth-generation Verano EPYC with LPDDR support, targets the cost-per-watt efficiency curve that increasingly defines competitive AI training infrastructure.In Europe, the European Commission is in active discussions with Imec and Germany's SPRIND on a 1nm fabrication facility by 2035 under the Chips Act 2.0 framework. The ambition is clear; funding mechanisms, location, and private participation are all unresolved.Elsewhere, Radxa's Qualcomm-based Dragon Q8B and Q5E boards signal a push to build a commodity edge AI ecosystem around Snapdragon silicon — though Windows 11 on ARM overhead remains a real friction point. And AI data centres are beginning to function as genuine grid infrastructure, with rack densities above 140kW driving mandatory liquid cooling and new potential for grid-balancing services.A YesWee production, built using AI technology.This episode includes AI-generated content.