TSMC printed its fifth consecutive record quarter on July 16. Revenue hit $40.2 billion, up 36% year-over-year. Net profit surged 77% — a number so large it barely registers. The business press covered it the usual way: AI boom, chip demand, record profits, repeat. They missed the sentence that matters to developers.
In the earnings call, CEO C.C. Wei flagged something worth bookmarking: agentic AI is now creating a second, separate wave of CPU demand inside data centers — on top of the GPU demand that already sold out TSMC’s 2nm capacity through 2028. “No matter what CPU approach is taken,” Wei said, “whether it’s x86, Arm-based, or RISC-V architecture, they are almost all TSMC’s customers.” This is not an incremental update. It is a structural shift in what your infrastructure actually costs to run.
What 66% HPC Revenue Actually Means
High-Performance Computing — the segment that includes AI accelerators and server CPUs — now accounts for 66% of TSMC’s wafer revenue, rising 20% in a single quarter. Smartphones are down to 22%. This is a reversal of the historical revenue mix that took less than 18 months to complete.
The reason is not just that more AI chips are being made. Agentic AI workloads — the kind that plan, retrieve context, call tools, reflect on results, and self-correct — require orchestration CPUs running in parallel with accelerator GPUs. The chatbot era was GPU-dominated. The agent era adds CPUs back into the equation. TSMC’s CEO named x86, Arm, and RISC-V in the same breath. Intel, AMD, Ampere, and SiFive all compete for the same foundry capacity your GPU vendor does. The demand pressure is now multi-architecture.
Three Constraints, All Sold Out
Before planning any major agentic infrastructure investment, understand the actual supply picture. According to a detailed capacity analysis, three constraints are simultaneously sold out:
- 2nm process capacity: Booked through 2028. Apple consumes roughly half for iPhone and Mac chips. Nvidia, AMD, and Qualcomm divide most of the rest. New customer allocation does not exist.
- CoWoS advanced packaging: This technology stacks HBM memory onto AI accelerators. Nvidia holds approximately 70% of CoWoS-L capacity. TSMC is scaling from 35,000 wafers per month (late 2024) to 130,000 by end of 2026 — and it is still fully sold out through 2027.
- HBM memory: Samsung and SK Hynix have signaled 15–20% contract repricing. If you do not have a long-term hyperscaler supply agreement, you are paying more.
This triple constraint is structurally harder to unwind than 2024’s H100 backlog. That was one layer. This is three simultaneously.
Your Agentic Cost Model Is Probably Wrong
The chip supply crunch shows up in API pricing, cloud GPU availability, and inference latency. But the bigger problem is internal: most teams building agentic systems have not modeled their actual compute costs.
A simple chatbot query costs roughly $0.0001 to $0.001. A multi-step agentic task — one that plans, retrieves, calls tools, reflects, and self-corrects — costs $0.10 to $1.00 per completion. That is a 100x to 1,000x multiplier. Subagent fan-out, where a coordinator agent spawns parallel sub-tasks, can generate bills in the thousands. One documented enterprise incident produced a $47,000 charge from a single agentic run. Uber’s CTO put it bluntly: “The budget I thought I would need is blown away already. The entire annual AI budget was gone.” The average enterprise AI budget grew from $1.2 million in 2024 to $7 million in 2026, with Fortune 500 companies reporting monthly inference bills in the tens of millions.
What to Do Before 2027
TSMC’s Arizona fabs will help — eventually. Fab 21 Phase 2 begins 3nm production in the second half of 2027. Phase 3 (2nm) follows in 2028–2029. The $265 billion total Arizona commitment is real. It is also a 2028 problem. The infrastructure economics you face today will not improve materially in the next 18 months. The work is on your end:
- Set hard subagent fan-out limits. Per-workflow caps, not just per-request limits. Most orchestration frameworks support this. Most teams have not configured it.
- Route by task complexity. A DeepSeek R1 plus Claude Sonnet hybrid delivers comparable coding accuracy at 14x less cost than OpenAI o1 exclusively. The savings fund your infrastructure team.
- Budget for CPU costs explicitly. API pricing obscures the orchestration layer. When you move to self-hosted inference, CPU burn for the planning and routing layer surprises most teams.
- Know your self-hosting inflection point. Private deployment becomes economically attractive at around one billion monthly tokens for most workloads. Calculate where you are on that curve.
TSMC’s record quarter is a signal, not a reassurance. When the company that manufactures chips for every major AI vendor tells you demand is accelerating because of the specific workload category you are building, and that supply will not catch up until 2027 at the earliest, the right response is to update your cost model. Most teams have not done that yet.
