Summary

TSMC’s pricing restraint is less a mystery of bargaining power than a legacy of the customer-first model that built its dominant position. It won trust by guaranteeing capacity, not competing with clients or taking their IP, and running paid-off fabs for decades; Thompson thinks that habit left 5 nm and 3 nm underpriced just as leading-edge fabs became shorter-lived and cost “well into the 30 billions.”
The bigger investor issue is capacity: TSMC plans $52–56 billion of capex this year versus $41 billion in 2025 and roughly $30 billion in preceding years, yet cost inflation means supply will grow less than the headline increase. Today’s investment decisions shape 2028–29 output; TSMC rationally fears a 2028 AI-bubble break leaving idle fabs, but its restraint may leave customers undersupplied.
TSMC is not eliminating risk; it is transferring it downstream as permanently foregone revenue for NVIDIA, Microsoft, Google and other AI buyers. Thompson’s “TSMC brake” is the supply ceiling behind CEOs saying they could sell more with more capacity: choosing the safest foundry may actually load greater fab risk onto customers.
The call to action is for hyperscalers and model labs to make Intel and Samsung credible second sources now, despite three-to-four-year porting cycles and near-term execution risk. Competition would make TSMC fear losing sticky, multi-year customers more than distant overcapacity; simply flying to Taiwan to ask it to “please invest more” enables the brake.
For AI buyers, the desirable end state is deliberate foundry overbuilding, because the downside is excess capacity and “super cheap chips” while the upside is enough supply to monetize demand. Thompson calls current buyers “chickens” for avoiding Intel and Samsung; if they wait, too little capacity could itself help burst the AI bubble in 2028–29.
The supply-chain concentration is geopolitical as well as financial: Sharp objects that Dario Amodei likened China-bound AI chips to nuclear weapons to the absence of any mention of Taiwan. If the “nuclear weapons plant” sits roughly 80 miles off China’s coast, alternative foundries hedge both undersupply and the war scenario Sharp raises.

Deep dive

1. TSMC’s neutral-foundry promise became a geopolitical choke point

Sharp’s opening rant targets an omission: whenever Dario Amodei compares supplying China with giving nuclear weapons to North Korea, Sharp searches for “Taiwan” and finds nothing. If AI chips are that strategic, he asks, what does it mean that their essential “nuclear weapons plant” is roughly 80 miles from China?
TSMC began about 40 years ago with process technology from Philips and little else. It trailed the leading edge by perhaps five generations, so its proposition was reliability: customers received guaranteed capacity rather than being displaced whenever an integrated producer such as Intel or Texas Instruments needed its own factories.
Neutrality supplied the second moat. Because TSMC made no competing chips, it could promise, “We’re not going to crowd you out,” and would not appropriate customer IP; Sharp notes that Samsung could separate foundry and product operations, but Apple would still face a fundamental conflict when manufacturing through a smartphone rival.
Sharp wonders whether a Taiwan disruption would merely leave consumers using iPhone 13-level chips. Thompson reverses the premise: attention fixates on leading-edge processors, but the harder replacement problem is the multitude of legacy chips for which the US has little capacity beyond whatever GlobalFoundries might supply.

2. Leading-edge economics broke the run-it-forever pricing model

TSMC historically depreciated a fab over perhaps five years but operated it far longer; some late-1990s facilities still produce ancient, standardized parts for pennies after their capital costs have been recovered. That longevity joined a “customer-centric, customer-first mindset” to an ingrained low-cost culture.
The model weakened at a turning point around 2014 or 2015, when chips became more expensive. Some buyers stopped at 28 nm—where Thompson says China invested heavily—because performance gains no longer justified the premium; 7 nm, described in the discussion as the first EUV node, became somewhat stranded as leading-edge customers moved onward.
TSMC’s decision to rework some 5 nm capacity for 3 nm signaled that advanced fabs were not merely costlier but shorter-lived. The company had to become “like Intel” in a positive sense: capture value upfront rather than rely on decades of trailing revenue. Thompson thinks 5 nm and 3 nm were underpriced precisely when there was “no alternative.”
Thompson’s “Ben conjecture,” explicitly “not reporting,” is that insufficient pricing might help explain former CEO and chairman Mark Liu’s abrupt retirement, perhaps with Morris Chang’s fingerprints involved. The contrast is Chang’s 2008 return: amid recession and retrenchment, he called the downturn TSMC’s biggest opportunity and invested into the smartphone era.

3. Today’s capex sets a hard ceiling on 2028–29 AI supply

After ChatGPT happened, around 2022, TSMC had to estimate demand for 2025 and 2026. Thompson argues that its conservative 2022–24 spending decisions now explain why capacity falls far short of AI demand: semiconductor commitments take years to become working supply.
Planned capex rises to $52–56 billion this year from $41 billion in 2025 and roughly $30 billion in earlier years. Sharp initially reads that as nearly a doubling; Thompson corrects him to roughly 25% year over year, while stressing that pricier equipment makes the relationship between dollars and added capacity worse than linear.
That spending is principally a decision about 2028 and 2029, not 2026. TSMC’s fear is reasonable: if the AI bubble breaks in 2028, several years of investment could leave expensive equipment without customers. Because the business is so capital-intensive, Thompson says the outcomes can run from extraordinary profitability to, at the extreme, “you’re going bankrupt.”

4. TSMC’s caution converts foundry risk into customer revenue loss

Thompson’s central mechanism is that “risk doesn’t disappear from the system.” By limiting overbuild risk on its own balance sheet, TSMC creates foregone revenue for NVIDIA, Microsoft, Google and others; when CEOs say, “If we had more capacity, we would’ve sold more,” those permanently lost sales are downstream of insufficient foundry capacity.
This is the “TSMC brake”: one supplier gates the AI infrastructure build-out and moderates potential bubble conditions. Customers believe they are de-risking by choosing the best process and most dependable service, but constrained output means they are “loading on fab risk to themselves.”
Sharp’s formulation, which Thompson accepts, is that TSMC would rather choose something like a 67% increase over earlier spending than a 167% increase, sacrificing possible profit to limit tremendous downside risk. Thompson’s qualification is decisive: “TSMC is doing what is right for TSMC,” and can do so because it lacks competition.
Asked whether dependence on NVIDIA, Apple and AMD creates customer-concentration risk, Thompson says the immediate problem is the opposite: TSMC has too many customers and insufficient capacity. It must avoid prices high enough to prompt departures while supplying enough leading-edge chips that customers do not leave involuntarily.

5. AI buyers must manufacture the competition TSMC lacks

NVIDIA previously maintained leverage by splitting Ampere-generation production—gaming chips at Samsung and server chips at TSMC. Today, TSMC remains best; Intel 14A “maybe” looks promising, but committing now may not produce chips for three or four years, and mapping a design to another process is never trivial.
A credible Intel or Samsung would change TSMC’s governing fear from unused capacity five years out to losing customers today. Foundry switches are multi-year commitments and difficult to reverse, so prospective customer defections would pressure TSMC to invest more while the new suppliers also added capacity.
Thompson calls the approach of flying to Taiwan and begging TSMC to invest more “totally wrong”; he says Sam Altman should explore Intel instead. Limited supply lets TSMC choose winners, leaving customers afraid to antagonize it, but begging merely enables the brake. That is also his response to OpenAI’s repeated “we don’t have enough compute” complaint: help create the missing alternative.
Buyers should want every foundry overbuilding: the downside is “super cheap chips,” lower capital costs and ample volume. Thompson says self-described capitalists must “sack up”; otherwise the 2028–29 opportunity might be killed by capacity itself. Sharp adds war as another diversification case, while Thompson calls long-horizon hardware planning “a new muscle for Silicon Valley.”

Why Every Brain Metaphor in History Has Been Wrong [SPECIAL EDITION]Inside America’s AI Strategy: Infrastructure, Regulation, and Global Competition