TSMC's June 2026 revenue came in at NT$442.68 billion — up 67.9% year-over-year, an all-time record. H1 2026 totaled NT$2,404.48 billion, up 35.6% versus the first half of 2025. Q2 guidance stands at $39.0–40.2 billion; analyst consensus is $40.04 billion versus Q2 2025's $30.07 billion. The company has beaten estimates for eight consecutive quarters. Full earnings are tomorrow, July 16.

The headline number is impressive. The thing the headline number implies is more interesting.

N3 — the manufacturing node behind every leading AI GPU and CPU this cycle, including NVIDIA's H200 and B200 — is sold out through year-end. CoWoS, the advanced chip-packaging process that bonds high-bandwidth memory to GPU dies and is what makes frontier AI inference hardware viable at scale, is also sold out through year-end. Revenue is a record because demand is a record. Capacity is not keeping up. You cannot solve either problem by spending more money in the near term.

That constraint is why Anthropic's move this month matters.

The chip race beneath the model race

On July 2, TechCrunch and Bloomberg reported that Anthropic is in early-stage discussions with Samsung Electronics about a custom AI inference chip. Target process: Samsung's 2nm foundry. Goal: inference, not training.

The caveats on this are real. Anthropic hasn't begun detailed chip design, testing, or manufacturing specifications. The company is still determining what the chip should do and how it would fit into a server. Custom chip timelines run 18 to 24 months from concept to production volume. First chips, at best: late 2027.

So why do this now? The Colossus compute deal tells you. Anthropic has been paying xAI $1.25 billion per month to run Claude on Elon Musk's data centers. That is an extraordinary number. It's also what happens when your primary product depends on compute you can't control the cost of and can't build fast enough to meet demand.

Custom chips are the exit from that dependency. The exit takes two years to build.

Why Samsung? Samsung's previous talks with OpenAI about a custom ARM-based inference chip stalled in early June 2026. So Samsung has available design capacity, a team already oriented toward AI inference silicon, and a motivation to close a deal Anthropic can actually commit to. Samsung's 2nm process also competes with TSMC's N2 — but it's not TSMC, which is the point. You can get on Samsung's roadmap. You cannot currently get meaningful new allocation on TSMC's N3.

Source spread

• CNBC — TSMC June revenue rises 68% ahead of second quarter — builder. Primary source for the June revenue figure and YoY growth; also covers the N3 and CoWoS sold-out context.
• TechTimes — TSMC sets all-time revenue record — hype. Strong on what the N3 sellout means for the next two quarters of AI hardware supply.
• TechCrunch — Anthropic is discussing a custom chip with Samsung — builder. Original report on the talks; good on the early-stage caveats.
• Bloomberg — Anthropic in talks with Samsung for custom AI chip — builder. Independently verified; context on Samsung's stalled OpenAI talks.
• Forbes — TSMC's second quarter will test whether the AI buildout has a ceiling — skeptic. Makes the case that N3 sellout and guidance may be near the demand ceiling rather than evidence of unconstrained upside.

Pros & cons

What's real:

• TSMC's revenue numbers are not speculative — the June 6-K filing is publicly available. NT$442.68 billion at 67.9% YoY is confirmed. The constraint story (N3 and CoWoS sold out) has been independently verified across multiple analyst reports. This is not AI demand narrative; it's actual fabrication capacity against actual order books.
• Anthropic's Samsung move is the right strategic logic. Distributing foundry risk away from pure TSMC dependency, targeting inference-specific silicon, approaching Samsung while the company has available capacity — all of that is sound. Custom silicon for inference is where every major lab has to go eventually if they want to control their cost curves.
• Samsung's foundry talks with OpenAI stalled in June, which means Samsung has a motivated design team, cleared capacity for new discussions, and an obvious reason to move fast with Anthropic. Timing matters in chip partnerships. Anthropic showed up at a good moment.

What deserves a side-eye:

• "In talks" is not "doing." OpenAI announced Jalapeño on June 25 with Broadcom as a design partner — a company with deep ASIC expertise. Anthropic has discussions with Samsung. Those are different starting points. Custom chip development requires sustained organizational commitment across multiple years and a design team with specific expertise Anthropic hasn't publicly demonstrated it has.
• Samsung's advanced node yield rates have historically trailed TSMC's on comparable process generations. The 2nm may narrow that gap; it's unproven at volume. The tradeoff is access (Samsung can take the design; TSMC N3 cannot) versus yield risk (Samsung may deliver more defective chips per wafer in early production). That's a real engineering tradeoff, not just a supply chain preference.
• TSMC's earnings call tomorrow will have updated full-year guidance. If TSMC raises its revenue outlook substantially, demand is continuing to outrun capacity and the AI capex cycle has legs. If guidance holds or softens, the Forbes-skeptic case — that the AI buildout is approaching a spending ceiling — gets more credible. The call is the data point that either confirms or complicates everything I've written here.

What builders need to know

• TSMC earnings are tomorrow, July 16. The full Q2 numbers plus updated full-year guidance land on that call. If you're making infrastructure decisions or vendor commitments based on AI compute availability, the guidance raise (or lack of one) is the number to watch — not the revenue beat, which is already priced in.
• N3 and CoWoS sold out through year-end means GPU supply is physically capped. You cannot expand frontier inference capacity by spending more if the wafers don't exist. If your roadmap requires meaningfully more GPU-hours before Q1 2027, you're either already on an allocation list or you're waiting.
• Anthropic custom chip timeline: late 2027 at earliest. Don't let this announcement change your API cost modeling for the next 18 months. The Sonnet 5 and Opus 4.8 pricing you're budgeting against is the pricing you plan around through 2026 and into 2027.
• Samsung 2nm vs TSMC 3nm: if both Jalapeño (OpenAI) and Anthropic's chip land on schedule, OpenAI's is at a more established foundry with a more mature design partner. Anthropic's is at an available foundry with an unproven yield baseline on the target process. That's not a knock on Anthropic's strategy — it's the tradeoff they made to get access at all.
• Google is already past this problem. TPU v8 (both 8t for training and 8i for inference) is in production now. The TSMC constraint still affects them at the margin, but they have dedicated capacity through their cloud infrastructure agreement. The strategic position of having your own silicon is paying off right now in ways the public numbers don't fully capture.

Further reading

• CNBC — TSMC June revenue rises 68%, Q2 beats guidance — primary revenue source
• TechTimes — TSMC sets all-time revenue record: AI demand has rewritten its calendar — N3 and CoWoS sold-out context
• TechTimes — TSMC Q2 earnings July 16: three CoWoS signals that test AI's spending ceiling — CoWoS capacity analysis
• Forbes — TSMC's second quarter will test whether the AI buildout has a ceiling — demand ceiling skeptic case
• TechCrunch — Anthropic is discussing a new custom chip with Samsung — original report
• Bloomberg — Anthropic in talks with Samsung for custom AI chip — independent confirmation and Samsung-OpenAI stall context
• TechTimes — Anthropic talks Samsung, aiming at 2nm process — chip design timeline details