Steve Hou
Head of Research at Silicon Data, AI Compute Markets
Chip testing may be the new memory: a thinly occupied, high-moat niche unknown to most investors. With far more chips everywhere, including GPUs and HBM, they cannot be installed without robust testing, which favors names such as Camtek, Aehr Test Systems, Teradyne and FormFactor. Higher valuations and slower earnings growth are no objection, since memory also looked that way many hundreds of percent ago. x.com x.com x.com x.comÂ
Power semiconductors and testers stand to gain as AI racks move to 800-volt DC, which raises power-semiconductor content per rack substantially. Per @meggmcnulty, doubling the voltage halves the current and cuts copper resistive losses by about 75%, with SiC and GaN suppliers and rack-level DC-DC converter makers gaining share. Aehr has outperformed as a small stock that inflected, and the group has re-rated upward, its significance only now rising as compute and energy demand surge. x.com x.comÂ
Aehr Test Systems stands to benefit because testing is a bottleneck to scaling AI chips. Testing demand is about to surge. Investors did not question its price-to-sales ratio at 120, despite the current P/S objection. x.com x.comÂ
FormFactor and Aehr Test Systems are favored because testing is the one non-optional part of the AI buildout. A single bad die ruins the whole package, so pricier chips mean more testing. FormFactor's probe cards touch every wafer; Q2 set records above a $1B run rate with gross margin over 50%, HBM4 is two-thirds of DRAM revenue and a new GPU program ships this half. Aehr guides fiscal 2027 revenue up 160-200% to $130-150M on a record backlog, plus a $22M AI follow-on order. x.comÂ
Quantum computing's suppliers get paid whoever's machine wins, as steel mills did while most railroads went bankrupt. Chips, test benches, cryogenic freezers, inspection, lasers, vacuum gear and light must be bought regardless. Washington signed nine letters of intent worth about $2.013 billion before any machine was built, half to the one recipient with a working foundry. Only one name is in its buy zone; the rest await lower prices. Two flat quarters of cryogenic and control orders ends it. … Show more
As AI makes memory the bottleneck, value is shifting from compute into the memory system, and HBM test and process control, which validates every HBM stack before delivery, sits outside the already-priced HBM narrative. HBM is growing from $35 billion in 2025 to an estimated $100 billion in 2028, with capacity sold out through 2026. But HBM makers and TSMC's CoWoS are now 2-3x return cases, not 10x; the opportunity lies in strong nodes with pricing room left. x.comÂ
FormFactor is evolving from an HBM probe-card play into a broader AI test platform as Nvidia emerges as a new logic-test market. Per Deutsche Bank, becoming a second source for Nvidia GPU probe cards at TSMC could reach $200M+ revenue by 2028, alongside about $400M of HBM revenue and a roughly $200M custom ASIC opportunity. Micron's print shows HBM demand exceeding supply, and taller, denser stacks make wafer-level testing more valuable. Valuation is the main risk after a 274% one-year rise. … Show more
Semiconductors could capture scarce spot compute's value if its forward curve holds. Compute trades in deep contango, while chip stocks price in little of that value; near-term moves are unpredictable. That makes the sector very bullish. The earlier lab-funding risk looks less acute: Anthropic should hold nearly $200B cash after its IPO and is operating-cash-flow profitable. Even with 2029 capex cuts, chips look cheap if future buyers do not default. x.com x.com x.com x.com x.comÂ
FormFactor sits in test and packaging, the best corner of semiconductors this year, where much quiet money is being made in the layer that makes compute functional. It makes the probe cards that test chips and ranks first globally in test subsystems. Last quarter set a revenue record, up 32%, with gross margin up 510 basis points to 49% and EPS beating by 33%. DRAM probe card revenue hit a record on HBM, a product sold out through 2030. x.comÂ
SK hynix's record earnings and five-year volume and spec agreements make its valuation look too low. Quarterly revenue reached KRW104.62T, up 32%, and operating profit hit $63B against a $785B market cap, below AMD's and less than half Tesla's. Its planned KRW100T (~$75B) shareholder return would be one of the world's largest, signaling conviction and unusual demand visibility. AI-inference 3D DRAM is coming by 2030; a reasonable valuation is all that is sought. x.com x.com x.com x.com x.com … Show more
As LLMs commoditize, chip programs will multiply, favoring vendors paid per attempt: CEVA, Aehr Test Systems, inTEST, PDF Solutions and GSI Technology. Stable, cost-driven inference makes hardwired custom chips worth building, and every program pays for IP, yield and test, success or not. CEVA earns per-unit IP royalties, Aehr and inTEST scale with SKU count, PDF Solutions owns cross-fab yield data, and GSI is a risky memory-wall option. Ongoing architecture churn would make this early. x.com … Show more
Semiconductor leaders look better placed than high-beta stocks with poor company models. TSMC, ASML, Nvidia and memory should lead the sector and belong in the buy basket, with Broadcom not far behind. The CPU basket is only a hold: much of its upside was priced in during Intel and AMD's run. x.comÂ
