The Compute 100 — Friday May 22, 2026 — #5

Micron Technology @MicronTech · May 21, 2026

"We're proud to announce a milestone moment for U.S. memory manufacturing. Micron has officially started manufacturing 1α DRAM at our Manassas, VA fab. As the only U.S. manufacturer of memory, Micron is strengthening America's domestic memory supply with the world's most advanced DDR4 technology — advancing our more than $200B U.S. manufacturing and R&D investment plans."

Yes, this is DDR4 — not HBM4. But we'd encourage readers not to gloss over the $200B number. Micron is the only U.S. memory manufacturer, full stop. The same industrial machine behind this Virginia announcement is what HBM4 supply for Rubin depends on. Keep an eye on Boise fab yield reports alongside this ramp.

AMD $449.59 ▲ 0.4% Mkt Cap: $728B

AMD put $10B+ into Taiwan ecosystem commitments this week, anchored around Helios — their rack-scale platform pairing 6th Gen EPYC "Venice" CPUs on TSMC 2nm with Instinct MI450X GPUs — targeting multi-gigawatt deployments in H2 2026. They also qualified the industry's first panel-based 2.5D EFB (Elevated Fanout Bridge) interconnect with ASE and SPIL. Panel-based packaging matters at rack scale: higher bandwidth density, better power efficiency, and the kind of supply chain optionality that becomes important when you're shipping at volume. ODMs Sanmina, Wiwynn, Wistron, and Inventec are building the systems. The +0.4% stock reaction is telling — this is a confirmation story, not a surprise. What we're watching is execution: whether AMD closes the rack-scale gap against NVL72 in actual deployments by year-end. The EFB qualification is a concrete step in the right direction.

ARM Holdings $298.23 ▲ 16.2% Mkt Cap: $313B

ARM hit a 52-week high of $298.70, up ~34% over two days, as post-earnings analyst upgrades stacked up: Jefferies to $290, TD Cowen to $265, KeyBanc to $300, Bernstein initiating at $300 Outperform. The underlying story is simple. Data center royalties doubled YoY in Q4 FY2026 — the second consecutive quarter of doubling — and ARM guided for another double in FY2027. Nvidia's print accelerated the move: Nvidia's networking segment tripling YoY is exactly the scale-up buildout that drives ARM's Armv9 royalty base. We've said before that ARM looks more like a royalty stream on AI infrastructure than a traditional chip company. These numbers are why.

Modal fundraise

$60M ARR in September. $300M ARR today. Eight months. That's the Modal story. The company raised $355M in a Series C at a $4.65B valuation — a 4.2x step-up from its $1.1B Series B just eight months ago — led by General Catalyst and Redpoint, with Menlo, Bain Capital Ventures, and Accel joining. The 5x ARR growth is driven by Modal's infra for AI agents. As they said - "In the last six months, it's become clear: agents are going to be everywhere, and they're far more powerful when they have a runtime to operate in". Modal blog →

Anthropic partnership

Anthropic is reportedly in early talks to run Claude inference on Microsoft's Maia 200 chips — per The Information (Bloomberg, CNBC, May 21). Nothing finalized. But the strategic logic is hard to argue with. Maia 200 on TSMC 3nm, with 30%+ tokens-per-dollar vs. existing GPU fleet per Microsoft's April earnings. If this closes, Microsoft gets its first major third-party inference customer for custom silicon — a meaningful signal in the AWS Trainium vs. Google TPU vs. custom silicon race. Anthropic has already committed $30B in Azure spend; running inference on Maia turns that commitment into a flywheel for Microsoft's silicon economics. We'll be watching closely.

Nvidia CFO Colette Kress with Tae Kim

Tae Kim's interview with CFO Colette Kress: On supply chain: Nvidia co-designs products with all three memory suppliers through multi-decade relationships that, she says, no one else in the industry replicates. On Vera Rubin: it's on track, despite reports of thermal issues. On agentic AI: she's citing immediate, measurable customer ROI as the engine of demand acceleration into 2027. And on capital returns: "$1 trillion returned to shareholders is on the horizon. Do the math."

New paper: 2.8x training compute efficiency from feedback-conditioned pre-training

A new arXiv paper (2605.20285) is claiming 2.8x compute efficiency on dense transformer LLMs — same benchmark quality at less than half the compute. The method is called Introspective X Training (IXT): it uses a "thinking reward model" to annotate training data with critique-based feedback, then routes those quality signals back to pre-training. Models tested: 7.5–12B parameters, trained to 18 trillion tokens.