The Lyceum: Semiconductor Weekly — May 06, 2026

The AI build-out is no longer the story — it's the gravitational field bending everything else. AMD's data center business grew 57% year-over-year, Infineon's AI power chips are supporting results amid a stalled EV business, and South Korea's KOSPI broke 7,000 for the first time in history amid what was effectively a two-stock rally. Underneath all of it, a quieter inversion: NAND contract prices are now climbing faster than DRAM, and the eMMC chips that go into every cheap laptop and car infotainment unit are about to become the squeeze nobody planned for.

• Zyphra Cloud Platform on AMD Instinct MI355X (Zyphra + AMD): Full-stack AI platform announced May 4, bringing Zyphra's research models to production on AMD's latest Instinct accelerators.
• SPEC CPU2026 Benchmark Suite (SPEC Consortium): Released May 5, the first major refresh in nine years — resets the public baseline that vendors tune cache, branch prediction, and memory subsystems against.
• Tower Semiconductor SiGe Beamforming ICs (Tower + Axiro): New family of silicon-germanium beamforming ICs ramping to volume production for U.S. defense radar and SATCOM phased arrays.
• SK Hynix HBM3E Modules for Nvidia (SK Hynix): Mass-production volume shipping into Nvidia's next AI platform, per the company's previously announced ramp.

If you've been wondering whether the AI infrastructure build-out is real or just Nvidia's story, AMD just handed you the answer.

AMD reported Q1 2026 earnings of $1.37 per share on revenue of $10.25 billion, beating consensus of $1.28 and $9.89 billion. Data center revenue alone hit $5.78 billion, up 57% year-over-year — the segment that houses EPYC server CPUs and Instinct MI-series GPUs competing directly with Nvidia for AI training and inference. AMD guided Q2 to roughly $11.2 billion, a beat-and-raise that drove after-hours gains.

What changes if AMD executes? The company is preparing Helios, its first rack-scale system pairing its GPUs and CPUs in a configuration analogous to Nvidia's Vera Rubin NVL72 racks. Hyperscalers buy compute at the rack level, not the chip level — Helios is AMD's bid to compete on Nvidia's actual playing field. Lisa Su told analysts AMD is now "a core partner to the world's largest AI infrastructure builders, with deep co-engineering relationships and multi-year visibility."

What failure looks like: Helios slips into 2027, or hyperscaler design wins narrow back to a few accounts. The signal to watch is TSMC's CoWoS-L (chip-on-wafer-on-substrate, large format) allocation in H2. If AMD's MI-series consumes meaningful incremental capacity there, Helios is real. If TSMC's packaging queue stays Nvidia-dominated, AMD's data center growth stalls at the chip level.

Infineon's quarter is a clean X-ray of two economies running at different speeds inside the same company.

The German chipmaker raised its fiscal 2026 revenue growth forecast to "significant" from "moderate," posting Q2 revenue of €3.812 billion. Inside the numbers: the Green Industrial Power segment — which makes power supplies for AI data centers — pulled €403 million against a €365 million consensus, with margin of 11.7% beating the 9.6% expected. Meanwhile, automotive margin came in at 18.1% versus 20.3% expected, dragged down by weak demand for high-voltage IGBTs (insulated-gate bipolar transistors — the power switches inside EV traction inverters that convert battery DC into the AC that drives the motor).

Infineon kept its FY26 AI power revenue target at €1.50 billion and guided €2.50 billion for FY27 — a 67% jump in a single fiscal year. That tells you how fast data center power density is climbing. Effective July 1, the company will reorganize from four segments to three (Automotive, Power Systems, Edge Systems), which reads like quiet preparation for a more aggressive capacity pivot toward AI power.

The signal to watch: Q3 IGBT order book. If orders don't recover, the EV slump is structural, not cyclical, and SiC and IGBT suppliers will have to rebalance toward AI applications faster than anyone planned. Infineon's design wins with Geely on battery management and central control units suggest China is the one EV market still absorbing new sockets.

The Semiconductor Industry Association reported worldwide chip sales hit $298.5 billion in Q1 2026, up 25% from Q4 2025. March alone reached $99.5 billion — up 79.2% year-over-year and 11.5% month-over-month.

A jump this large usually means multiple layers are firing at once: leading-edge logic, memory for AI servers, power semiconductors, networking silicon, and at least some recovery in the parts of the market that spent last year digesting inventory. This is no longer a narrow AI spike. It's starting to look like a broader capital cycle.

What changes if it holds: equipment makers see book-to-bill ratios push above 1.1, fab capex guidance gets revised upward at every major foundry, and the materials chain — gases, photoresists, substrates — starts hitting allocation. What failure looks like: Q2 prints flat or down, and the print becomes the cycle peak rather than its midpoint. The tell is TSMC's monthly revenue release in the first week of each month — sustained NT$420 billion+ months mean wafer starts are still accelerating.

The KOSPI broke 7,000 for the first time in its history, closing at 7,384, up 6.45% on the session. The rally was almost entirely two stocks. Samsung Electronics crossed a $1 trillion market capitalization, joining TSMC as the only Asian companies in that club, with local reporting placing the figure near 1,555 trillion won. SK Hynix hit fresh highs alongside it. [Source: Global Economic — Korean]

The catalyst is mechanical: Samsung and SK Hynix supply the HBM (high-bandwidth memory) stacks that go into every Nvidia, AMD, and Google AI accelerator. As AI infrastructure spending accelerates, both companies print record profits. South Korea's Q1 exports hit a record $219.9 billion, with semiconductor exports up sharply year-over-year — the macro version of the same story.

What matters beyond the stock charts is capex. A Samsung with a strong balance sheet is a Samsung that can fund HBM4 expansion and the 2nm GAA (gate-all-around — the next-generation transistor architecture that wraps the gate around the channel on all four sides) foundry ramp simultaneously. The signal to watch is Samsung's next capex guidance update — and whether ASML's High-NA EUV order book reflects a corresponding bump.

Last week's Hua Hong tool restrictions were the visible part. The structural change underneath them is now in motion.

The Bureau of Industry and Security has formally closed the Validated End-User (VEU) program, the license exemption that let foreign-owned fabs in China import most U.S.-origin equipment without individual licenses. Former VEU participants have 120 days from the rule's Federal Register publication to apply for export licenses. BIS says it intends to grant licenses to operate existing fabs — but not to expand capacity or upgrade technology.

That last clause is the one that matters. Samsung's Xi'an NAND facility, SK Hynix's Wuxi DRAM plant, and TSMC's Nanjing mature-node fab are now effectively frozen at their current process nodes. They can keep running. They cannot migrate to newer generations that would improve yield and cost. Meanwhile, per a Congressional Research Service analysis, domestic Chinese players including CXMT remain able to purchase U.S. equipment. The asymmetry — foreign-owned frozen, domestic Chinese tooling up — will reshape competitive share inside China's memory market on an 18-to-24-month lag.

The clock is running. Watch which foreign operators file the earliest license applications: that's the read on who believes the policy will hold.

For most of this memory cycle, DRAM was the headline. That's changed.

Per TrendForce's latest pricing survey, conventional DRAM contract prices will rise 58–63% quarter-over-quarter in Q2 2026, while NAND Flash will jump 70–75% — following Q1 DRAM increases of 90–95%. NAND's growth rate is now accelerating while DRAM's has begun to taper.

The collateral damage is in eMMC and UFS — the embedded flash that sits inside every mid-range smartphone, automotive infotainment unit, and industrial controller. TrendForce notes the eMMC/UFS segment faces the tightest supply gap of any NAND product because its process capacity overlaps with enterprise SSDs and offers significantly lower margins. It is the lowest-priority allocation. Meaningful capacity expansion is unlikely before late 2027.

What changes: automotive and industrial procurement teams who just finished the 2024–25 inventory correction get to start a new one. The PC side is already showing it — IDC expects global PC shipments to fall 11.3% in 2026 (per IDC's 2026 forecast) on memory cost. The signal to watch is spot pricing at DRAMeXchange and any commentary from Renesas, NXP, or Microchip on Q2 lead times for embedded flash.

Per AP News, TSMC has confirmed it will manufacture advanced AI semiconductors in Japan — a concrete, dateable expansion of leading-edge production outside Taiwan.

This doesn't replace the Taiwan capacity story; it complements it. Customers and policymakers now have another location to factor into sourcing strategies, and the move dovetails with U.S. and Japanese policy efforts to diversify advanced silicon. The corroborating signal: Lam Research disclosed that estimated future revenue from shipments to customers in Japan — tools shipped but not yet revenue-recognized — rose to roughly $434 million as of March 29, up from $226 million at end-of-December. That's not an order number; it's tools physically arriving on Japanese fab floors faster than they're being qualified.

What success looks like: a second non-Taiwan node-leading site at scale by 2028. What failure looks like: Japan becomes a trailing-edge satellite while the real 2nm and 1.4nm capacity stays in Hsinchu and Arizona. Watch the equipment delivery cadence — it leads the production story by 12 to 18 months.

A trillion-dollar Samsung, a German chipmaker quietly rerouting capacity from EV inverters to AI rectifiers, and a forgotten flash chip in a Hyundai infotainment unit about to become the supply-chain villain of 2026. The leading edge gets the press release; the eMMC inside your kid's tablet gets the allocation letter. Until next week.

Forward this to the procurement lead who's still treating embedded flash as a commodity — they have about a quarter to update their model.