Back to Insights
AI

TSMC's Blowout Quarter Is the Other Half of IBM's Bad One

TSMC beat Q2 guidance, raised its 2026 outlook past 40%, and added $100B for Arizona — then the market sold the capex. IBM's crash is the other half.

S5 Labs Team July 17, 2026

TSMC reported second-quarter results on July 16 that beat its own guidance on revenue and on both margin lines, raised the full-year outlook to “slightly above 40%” revenue growth in dollar terms, lifted the 2026 capital budget to $60–64 billion, and committed another $100 billion to Arizona. The stock fell. The semiconductor complex fell harder — the VanEck Semiconductor ETF dropped 4%, Micron 6%, SK Hynix 11% in Seoul — and the Nasdaq 100 closed down roughly 1.6%, though TSMC only owns the chip leg of that: Alphabet fell 4.4% in the same session on a Bloomberg report that Gemini 3.5 Pro is months behind schedule, and the Iran conflict was in the tape too.

Two days earlier, IBM had told investors its own quarter came in at $17.2 billion against a consensus the wires put near $17.86 billion, and its shares fell 25.2% — about $67 billion of market value, in what CNBC and Forbes both called the worst single day in the company’s history. IBM’s explanation was that in the last weeks of June, clients pulled budget forward into hardware.

The same money is on both sides of those two days, and the story that makes is narrower and more useful than the version in circulation.

The Market Sold a Beat-and-Raise

Revenue landed at US$40.20 billion, up 33.7% year over year, at the exact top of the $39.0–40.2 billion range guided in April and above the roughly $39.94 billion consensus. Gross margin came in at 67.7% against 65.5–67.5% guidance, operating margin at 60.3% against 56.5–58.5% — 180 basis points clear of the top end. Two-nanometer shipments appeared at 3% of wafer revenue, the first time TSMC has broken the node out as a disclosed line, though N2 has been in volume production since Q4 2025, so that is a disclosure milestone rather than first revenue. The full-year outlook moved from “above 30%,” TSMC’s filed wording in April, to “slightly above 40%” — and that raise is in the written filing, on slide 10 of the earnings presentation exhibit, not merely in an executive’s mouth on a call.

One number needs a haircut before you use it. Net income grew 77.4%, to NT$706.56 billion, and that is not an operating result: non-operating income tripled, from NT$29.61 billion to NT$95.83 billion, and Investing.com reports roughly NT$63.2 billion of the increase came from the disposal and mark-to-market gain on the 8.1% stake in Vanguard International Semiconductor that TSMC announced selling in May — one non-recurring item worth about 95% of the entire year-over-year swing. Operating income grew 65.4%. Carry that one instead. The filing also footnotes that the quarter’s figures have not yet been approved by the board.

The capital budget is the one place where it matters which document you are reading, because no full-year capex figure appears anywhere in the Q2 filing. “We have decided to raise our full year 2026 capital budget to be between $60 billion and $64 billion” is CFO Wendell Huang on the earnings call, up from a $52–56 billion range TSMC said in April would land at the high end. The one number the market actually traded on exists only in a transcript.

It traded on it hard, and the filing explains why. Capex hit NT$496.00 billion in Q2, up from NT$350.76 billion in Q1, and free cash flow fell 17.5% quarter over quarter, to NT$287.36 billion, in a record-earnings quarter. First-half capex was roughly US$26.8 billion, so reaching $60–64 billion means spending about $33–37 billion in the second half — a 30–40% step up from a run rate that already compressed cash. Huang told the call the steep 2nm ramp will dilute gross margin three to four points in the second half, and Q3 guidance of 65–67% already sits below what Q2 delivered. This is the Meta pattern from April again: the capex line, not the results line, is what got sold.

Arizona Is a Third Tranche, Not a Re-Announcement

TSMC has now announced “$100 billion for Arizona” twice, sixteen months apart, and conflating the two is the easiest error available here. March 2025 expanded an original $65 billion commitment by $100 billion, to $165 billion. July 16 adds another $100 billion on top, which the Commerce Department release states in the same breath as the earlier one: total planned US investment is now $265 billion. The arithmetic settles it — a restatement would still total $165 billion. Four more fabs at 2nm and below, taking TSMC to 12 US facilities.

There is no timeline — Wei said the construction pace follows market demand — so none of this touches 2026 supply or 2026 pricing. Fabs announced today produce nothing for years, and a shortage being priced this quarter is not waiting for them. The investment is TSMC’s; the Commerce release wrapped around it is the administration’s.

What IBM Actually Said

The framing in circulation is that AI capex is eating enterprise software budgets. IBM’s own segment data contradicts it. Software revenue grew 5%. Infrastructure fell 7%, and that was mainframes lapping a z17 launch — inside the same segment, Distributed Infrastructure grew 37%, IBM’s best performance in reported history, with roughly $500 million of backlog, “With clients prioritizing infrastructure investments.” IBM never attributed the shift to AI at all, and Arvind Krishna assigned “the majority of our shortfall” to IBM’s own execution: “this quarter we faltered.”

What Krishna described is more specific than the headlines. “In the last few weeks of June, we saw clients shift their quarterly capex spend toward servers, storage, and memory purchases to secure supply-constrained infrastructure ahead of expected price increases.” That is a supply-and-price explanation: enterprises with real budgets front-ran a hardware price increase, and deals IBM expected to close slipped a quarter while they did it.

TSMC’s numbers say those buyers read it correctly. Wafer shipments rose 16.6% year over year; revenue rose 36.0%. The roughly 17-point gap is price and mix, not volume. The company that makes the constrained thing is charging more for it, says its own cost to make it is rising as 2nm ramps, and is committing $60–64 billion this year and $265 billion in the US to build more of it. Money did not leave the economy between July 14 and July 16 — it moved down the stack, toward the part that is scarce, which is why the “AI is killing software” reflex keeps mispricing weeks like this one.

If You’re Buying Either

If you have hardware coming in the next six to twelve months — an on-prem refresh, a NAS or backup expansion, workstations — the enterprises ahead of you in the queue already pulled their orders forward, and IBM’s letter is the receipt. RAM and SSD are the exposed line items. Get quotes early and hold vendors to them; last quarter’s price is not a planning assumption.

The subtler one lands on anyone costing out an AI feature. TSMC guided margin down while ramping 2nm because leading-edge silicon is getting more expensive to produce even for the company with the most pricing power in the industry, and the power and siting costs underneath it move the same way — New York just paused new large data centers over precisely that. If your model assumes per-token prices only ever fall, revisit it. Cheap open weights keep arriving; the compute they run on is not getting cheaper this year, and production cost models built on 2025’s price curve will drift.

The Constraint Underneath Both Days

One quarter is one quarter. IBM’s figures are preliminary and it reports properly on July 22; TSMC’s have not been approved by its own board; IBM blamed itself for most of the miss and is probably right to. None of this proves software is dying or the AI trade is cracking. It does establish that demand for leading-edge silicon is strong enough that TSMC raised its own full-year outlook by ten points in three months, tight enough that enterprise buyers are front-running price increases rather than waiting them out, and expensive enough to serve that the market punished the company reporting it. Every one of those is a reading of cost rather than appetite. Someone absorbs cost, and the further you sit from the fab, the more likely that someone is you.

Key Details

ItemDetail
CompanyTSMC (Taiwan Semiconductor Manufacturing Company)
QuarterQ2 2026, reported July 16, 2026 (figures not yet board-approved)
RevenueUS$40.20B (+33.7% YoY) / NT$1,270.38B (+36.0% YoY)
Gross Margin67.7% — above the 65.5–67.5% guided range
Operating Margin60.3% — 180bps above the 56.5–58.5% guided range
Net IncomeNT$706.56B (+77.4% YoY); operating income +65.4% is the clean figure
Node Mix2nm 3% (first broken out), 3nm 30%, 5nm 33%, 7nm 11%; 7nm and below = 77%
Q3 2026 GuidanceUS$44.6–45.8B; gross margin 65–67%, operating margin 56–58%
FY2026 Outlook”Slightly above 40%” revenue growth in USD — raised from “above 30%” (in the filing)
2026 CapexUS$60–64B, raised from $52–56B — earnings call only, not in the filing
Arizona+$100B third tranche; US total $265B; 4 new fabs, 12 US facilities; no timeline
IBM, July 14Preliminary $17.2B (+1%); Software +5%, Infrastructure -7%; shares -25.2%

Sources

Want to discuss this topic?

We'd love to hear about your specific challenges and how we might help.