TSMC Books the Boom, Valve Pays the Bill: The Chip Supercycle Has a Hardware Tax

What Happened

TSMC reported NT$1.27 trillion in second-quarter revenue, roughly $39.6 billion, up 36% from a year earlier. Bloomberg framed the number as the clearest sign yet that AI infrastructure spending has not cooled. The company then went further at earnings: capital expenditure guidance for 2026 rose to $60 billion to $64 billion, and the full-year revenue growth forecast moved from the 30% range to above 40%. At the top of the foundry stack, the supercycle looks unstoppable.

The same week, the view from the bottom sounded very different. Valve engineers Yazan Aldehayyat and Pierre-Loup Griffais, the people building the Steam Machine, told Bloomberg the memory market is, in Aldehayyat’s words, “honestly, still getting worse.” Retail prices trail the bulk supply market by three to six months, which means the sticker you see in a store today reflects a storm that already passed. The Steam Machine launched at over $1,000, and the pair expects prices to keep climbing into early 2027. Production is throttled too: “We’re basically building everything we can get our hands on. We’re limited by memory capacity, for sure.” DRAM makers, busy redirecting wafers to AI server memory, now sell month to month with no contracts. Griffais put the leverage bluntly: if a buyer says no, they never talk to you again.

Markets are pricing both stories at once. The Philadelphia Semiconductor Index is up 83% this year, yet at one point it fell more than 11% from its June record. In the week ending June 24, chip funds bled about $11 billion, the largest weekly outflow this century according to LSEG Lipper. Money keeps switching sides on whether the cycle holds.

What This Means for Founders

A supercycle delivers its profits and its costs to different addresses. The profits flow to TSMC and the memory makers selling HBM to hyperscalers. The costs flow to every team that buys DRAM and NAND to build a product. If Valve, with real cash and real volume, cannot lock in memory at a workable price, a seed-stage startup shipping robots, edge AI boxes, or IoT devices stands far behind it in line. The three-to-six-month retail lag is the trap to watch: cost a next-quarter build off today’s component quotes and your margin may be gone by launch.

The structure of buying has changed as well. Month to month, no contracts, take it or lose the relationship. In that market, purchasing power is everything, and small teams have none. Which turns memory-frugal engineering into a margin strategy: software that does the same job with less DRAM, SKUs split by capacity so price hikes flow into premium tiers, boards designed to accept whatever memory grade is available. The flip side is opportunity. Component procurement brokers, BOM risk simulation, refurbished memory marketplaces: when an input this universal gets scarce, selling relief becomes a business.

What You Can Do Now

If you ship hardware, recompute the memory share of your BOM this week. Run cost scenarios against delivery three to six months out, not spot prices, and check how long your supplier quotes actually hold. Lock volume early if you can afford it; if you cannot, split capacity tiers so the increase lands on the top SKU. Software founders are not exempt either. As customers’ hardware costs rise, anything that squeezes more work out of existing machines gets easier to sell. Price that pitch now.

See the Structure as a Map

The machinery behind this supercycle (the CUDA moat, the $40B equity strategy, the HBM lock-in) is laid out node by node in the DeepThought brief AI Platform Vertical Integration.