TSMC Bets $64B on AI Chips — But First, It Audits the Grid
By Meta, TechPulse USA
The most important company in the AI boom just told Wall Street it doesn't quite trust America's AI boom. TSMC — the foundry that makes nearly every serious AI chip on Earth — is sending its own people out to inspect US datacenter construction sites and power supplies before it commits $64 billion to new fabs. Orders aren't the problem. Electricity is.
That's the remarkable admission buried in the company's third-quarter earnings call on October 15, and it deserves more attention than the record numbers got. TSMC will spend $60 to $64 billion on capital expenditure this year and plans to spend more across 2026–2028 than it did in the previous three years combined. It is planning roughly 20 leading-edge manufacturing and advanced-packaging facilities in Taiwan and overseas. And before any of it happens, it wants its own eyes on whether America's datacenters actually get built — and plugged in.
The scouts are the story
The Register reported on October 15 that TSMC is dispatching personnel to check out US datacenter construction and power availability as it plans billions of dollars in additional chipmaking capacity. The detail surfaced on the earnings call, in the question-and-answer session, when one analyst asked the question everyone on Wall Street is tiptoeing around: could power shortages, datacenter construction delays, advanced-packaging constraints, or tight high-bandwidth memory supplies hold back projected growth?
Chairman and CEO C.C. Wei didn't dodge it. He described a capacity-planning process that treats the grid as a variable, not an assumption:
"We put every possibility we can, talking about the cloud service providers, capex, datacenter build-out, financial, power supply, and also other component supply, all into our consideration," Wei said on the call.
Translation: TSMC now audits the datacenter buildout the way a bank audits a borrower. It has sent teams to assess US data-center build-outs and power conditions — and according to Seeking Alpha analyst Julia Ostian, who follows the company closely, "the feedback is so far so good, and cloud providers are well-prepared." Even with the grid on the checklist, Wei's verdict was unambiguous: "The demand is very, very high," he said, adding that the gap between demand and TSMC's ability to supply remains substantial.
What Wei said about 2027 — the cautious bit
It would be a mistake to read this as pure swagger. Wei used the same call to sound a genuinely cautious note about the parts of the chip business that aren't AI — a warning that should interest anyone buying gadgets next year:
"Looking ahead, we continue to observe consumer and price-sensitive end market segments are being challenged due to the impact of rising component prices and macroeconomic uncertainties. As such, we are being prudent in our business planning going into 2027," he said.
Rising component prices, in plain English, means the memory crisis we've been documenting for weeks — Intel's CEO Lip-Bu Tan warned that memory prices are up five to seven times and that 2027 will be worse. Your phone is getting more expensive to build. AI accelerators are not.
On AI itself, Wei was all-in. Demand stays "extremely robust," with CPUs for agentic workloads and networking chips adding to the hunger for silicon. Management indicated that the compound annual growth rate for AI accelerators from 2024 to 2029 — previously forecast in the mid-to-high 50s — will "probably be higher." And he closed the thought with the line of the quarter:
"Therefore, semiconductors will remain essential in the AI mega trend. The value of semiconductors in the global supply chain is set to continue to increase in our view."
We covered the quarter itself — a record $23.6 billion profit that Wall Street still judged against perfection — when the numbers dropped. The call's real news was always the infrastructure caveat.
The $64B math
Put the spending in perspective. TSMC raised its 2026 capital expenditure range to $60–64 billion back in July, from $52–56 billion, and this month reaffirmed it. CFO Wendell Huang said spending from 2026 through 2028 should be significantly higher than the prior three years' total — and that if TSMC executes, revenue growth should outpace capex growth.
The physical plan behind the money: nearly 20 fabs under development worldwide, including 13 leading-edge and advanced-packaging facilities in Taiwan and five to six overseas. The 2-nanometer family, already 8% of wafer revenue in its early ramp, is expected to more than double in annual capacity in 2027. And A14 — the 1.4nm-class node Wei expects to become a bigger business than 2nm — is on track for risk production in 2027 and volume production in 2028, with customer tape-out activity running ahead of schedule.
Fourth-quarter guidance was a jaw-dropper on its own: $52.3–54.3 billion in revenue, roughly 58% year-on-year growth at the midpoint. But here's the thing about guidance like that — it assumes the datacenters exist. Hence the scouts.
Arizona, Texas, and the "run by TSMC" line
The scouting isn't only about datacenters — TSMC is also casing new fab locations, and the US is where the action is. Arizona remains the anchor: a $265 billion commitment (including the extra $100 billion announced in July), with the second fab's construction finished, tool installation underway, and volume production slated for the third quarter of 2027, slightly ahead of schedule.
But Arizona alone may not suffice. On the call, Wei confirmed TSMC has sent evaluation teams to Texas and is discussing Elon Musk's proposed Terafab project — a joint effort involving Tesla, SpaceX, and xAI — at the request of major American customers who want more onshore capacity. The Associated Press and CNA both reported that US government assistance is factoring into the deliberations.
The negotiation has a delicious wrinkle. Musk, posting on X earlier in October, said his companies would build and operate Terafab, with TSMC subleasing cleanroom space at most. Wei's answer on the investor call had the flat finality of a man who has spent four decades running fabs: "Either way, it will be possible, but the fab has to be run by TSMC." That gap is unresolved, and it's the kind of standoff worth watching — because whoever operates the fab controls the most advanced chipmaking on American soil.
The binding constraint is not what you think
Step back and appreciate the inversion. For forty years, the question in semiconductors was whether anyone would buy the chips. Now the world's biggest chipmaker is telling you that orders are fine, lithography is fine, and the thing keeping executives up at night is whether the electricians finish on time.
And power is only one item on Wei's checklist. The analyst's question bundled four constraints: power, datacenter construction delays, advanced packaging, and HBM supply. On packaging, TSMC is in production at 5x reticle sizes — claimed to be the only company doing it at scale — and working toward 9x and 14x, while developing new approaches alongside GlobalFoundries for backend capacity. On memory, the picture is uglier: our September reporting on the memory crisis showed prices multiplying, and HBM remains the choke point for every AI accelerator on the roadmap.
This is the same supply-side ceiling we wrote about in September, when Jefferies' satellite analysis found America is planning more AI datacenters than its chip supply can fill — roughly 11 gigawatts of US deployments this year, 16–18 next year, and a practical cap in the low 20s after overhead. TSMC's scouts are the chipmaker's version of the same audit: don't model the demand, go look at the construction sites.
Our take: the grid is the new lithography
Here's the honest read. Whether the AI boom continues will be decided less by Nvidia's next guidance call and more by whether the power actually shows up — at the right substation, on the right date, at a price anyone can afford. For a decade, America's grid planning has been botched, slow, and parochial; now it has to absorb the most capital-intensive technology buildout in human history, on AI's timetable, with the world's most important chipmaker sending inspectors to check the homework.
We checked The Register's dispatch account against the Motley Fool transcript of the October 15 call and three independent Q&A relays — and the "scouts" framing holds. Wei described the checks himself, in his own words, unprompted except by a direct question. This isn't a rumor. It's the foundry's operating procedure, stated on the record.
The comforting version of this story is that TSMC's teams will find the datacenters on schedule — Ostian says the feedback so far is good — and the $64 billion flows. The uncomfortable version is that TSMC felt the need to look in the first place. A company that plans in three-year, hundred-billion-dollar increments doesn't dispatch inspection teams for fun. It does it because, for the first time in its history, the binding constraint on selling more chips is something it can't manufacture: somebody else's infrastructure.
Update, Oct. 19: the tape disagreed — record profit, and the shares fell anyway
Taipei-listed TSMC shares closed 2% lower on Oct. 15 before the release, as the Taiex fell 3.41% in a chip selloff that also knocked more than 10% off Samsung and SK Hynix. Gross margin slipped 1.3 points to 66.4% on 2nm ramp costs — the quarter's one wrinkle. The street's verdict, not the tape's: at least four analysts raised price targets after the call. Freedom Broker's Egor Tolmachev kept a buy and lifted his target to 560 from 475, writing that "demand still runs ahead of supply"; Wedbush's Matt Bryson kept outperform, noting 13 leading-edge and advanced-packaging fabs in Taiwan plus five to six overseas. The line the relays buried: "We have no intention of leaving opportunities to others," Wei told the call — the same call where agentic-AI CPUs and networking chips were flagged as demand drivers on top of accelerators. CFO Wendell Huang said overseas-fab dilution runs near 2% this year and widens to 3–4% later as more US fabs ramp.
TSMC grid-audit FAQ
Why did TSMC stock fall after record earnings? Record NT$763.06B profit landed into a chip selloff — Taipei shares closed 2% down, the Taiex off 3.41% — and gross margin slipped 1.3 points on 2nm ramp costs. Analysts still raised price targets.
What is the Terafab standoff? Elon Musk's proposed joint fab (Tesla/SpaceX/xAI) at the request of major US customers; Musk says his companies will build and operate it with TSMC subleasing space; Wei answered flatly: "the fab has to be run by TSMC".
Is TSMC actually building in Texas? Evaluation teams have been sent to Texas; the AP and CNA report US government assistance is factoring into the deliberations.
Sources: The Register (Oct 15, 2026); Motley Fool Q3 2026 earnings call transcript; MarketBeat and LongBridge call highlights; WebProNews; TrendForce (Oct 15); Focus Taiwan/CNA (Oct 15); Zotpaper; Seeking Alpha via TradingView. Our September reporting on TSMC's Texas evaluation and the Jefferies datacenter-ceiling analysis, plus our Intel memory-price warning piece, supplied background context.
About TechPulse USA: TechPulse USA is an independent technology news publication covering the AI and semiconductor stories shaping America and the world. We research every story from multiple sources, verify every fact we print, and license every image honestly. Tip us: the story nobody else is covering is the one we want.
Comments
Post a Comment