If you pay for an AI subscription — ChatGPT Plus, Claude Pro, Gemini Advanced, whatever — you've probably noticed the price hasn't moved much in the last year. Twenty dollars a month has become the industry floor. Not the ceiling. The floor.

At the same time, API prices have collapsed. Frontier model output that cost $25 per million tokens a year ago now costs $4 to $6. Models are demonstrably better. Competition is fierce. And yet the subscription price your credit card sees every month is stubbornly the same.

TSMC's Q2 2026 earnings call, reported today, offers the clearest structural explanation for that gap.

The company reported NT$1,270.38 billion in revenue — approximately $40.2 billion, up 36% year-over-year. Gross margin came in at 67.7%. Operating margin at 60.3%. GAAP EPS hit $4.31 per ADR against a consensus estimate of $3.87. The company raised its full-year 2026 revenue growth guidance to greater than 30%.

A record quarter. By a lot.

TSMC matters here because TSMC makes the chips that run every model you've ever used. ChatGPT runs on NVIDIA GPUs. Claude runs on Google TPUs and NVIDIA hardware. Grok runs on NVIDIA. The GPUs and TPUs run on silicon fabbed almost exclusively at TSMC. There's no shortcut around this. If AI compute is expensive, TSMC's revenue line is part of why.

The number that matters most

HPC — High Performance Computing, which includes AI training and inference chips — accounted for 66% of TSMC's Q2 wafer revenue, up 20% quarter-over-quarter from 61% in Q1. A year ago, HPC was around 55% of TSMC revenue. TSMC started as a company that made chips for smartphones. It is now, structurally, an AI infrastructure company that happens to still make phone chips.

The other number: CoWoS. Advanced packaging — the technology that connects GPU memory to processors at the bandwidth AI workloads actually need — is sold out through year-end 2026. TSMC expects to produce approximately one million CoWoS wafers in 2026 — nearly three times the roughly 370,000 produced in 2024. Still not enough to clear the queue.

That sold-out status is why model API price cuts don't flow into your monthly subscription. NVIDIA makes the hardware. TSMC packages it. Both are capacity-constrained against actual demand. The AI labs buying compute at the frontier are doing so against a physical shortage, not a pricing competition. A company can't discount a subscription when its cost base is growing 36% per year at the physical layer.

N2 at 3% of wafer revenue is the forward-looking number. It's TSMC's newest process node, ramping at two sites, supported by smartphone, AI, and HPC demand. When N2 is at 20%, the cost curve for AI compute starts bending. That's probably a 2027 story.

Source spread

• TSMC Investor Relations — Q2 2026 Results — builder. Primary source: NT$1,270.38B revenue, NT$706.56B net income, 67.7% gross margin, $4.31 EPS.
• GuruFocus — TSM Q2 Earnings — builder. $40.2B revenue, $3.87 consensus EPS vs $4.31 actual.
• Investing.com — Earnings call transcript — builder. HPC at 66%, N2 at 3%, CoWoS sold out, >30% guidance, CC Wei statements.
• TechTimes — CoWoS capacity preview — skeptic. Pre-earnings analysis of CoWoS capacity signals and what they imply for AI infrastructure ceilings.

Pros & cons

What's real:

• AI compute demand is confirmed durable. A 36% year-over-year revenue gain at TSMC's scale doesn't happen on hype. It requires actual workloads — training runs, inference calls — running continuously at volume.
• HPC at 66% means TSMC has structurally realigned around AI. Capital allocation, capacity planning, process node roadmap: all of it now follows AI demand signals. That tends to create compounding advantages for the infrastructure layer.
• N2 at 3% is early, but it's ramping. When this number grows, the compute cost curve changes — for builders and eventually for everyday users.

What deserves a side-eye:

• "Greater than 30% growth" as full-year guidance is usefully vague. That range includes 31% and 49%. TSMC isn't telling you which.
• CoWoS being the binding constraint is a different kind of problem from wafer shortage. You can build more fabs. Advanced packaging expertise takes years to develop. The 1M-wafer projection for 2026 is extraordinary growth — and still sold out.
• The geographic concentration of all this in Taiwan, and the vendor concentration in a single company for nearly all frontier AI silicon, is a systemic risk that a record earnings call can't paper over.

What to do about it

For builders:

• CoWoS sold out through year-end means no meaningful capacity relief on H100/H200-class hardware before 2027. Build infrastructure plans that assume current constraint levels.
• N2 is at 3% of wafer revenue — early commercial ramp. N3/N5 remain the realistic options for any AI silicon design landing in 2026–2027.
• The >30% full-year guidance is the most important number to watch in Q3. If TSMC lifts it above 35%, the capex cycle is extending, not moderating.

For everyone else:

• The gap between falling API prices and your $20/month subscription is real and structural. AI labs compete on API pricing to win developers; consumer subscriptions are stickier because the physical cost base isn't declining.
• If you're deciding between free and paid tiers: some of what the paid tier buys is genuine compute priority during peak hours, not just model access. That's backed by real infrastructure cost, not marketing.
• The version of this that changes — lower subscription prices, genuinely — requires N2 volume at scale. That's a 2027 story at the earliest.

Further reading

• TSMC Investor Relations — Q2 2026 Results — full financial statements and press release
• GuruFocus — TSM Q2 Earnings — clean earnings summary with consensus comparison
• Investing.com — Earnings call transcript — CC Wei statements, HPC/CoWoS breakdown, guidance details
• TechTimes — Three CoWoS signals — pre-earnings context on what the packaging capacity signals actually mean