TSMC just made every advanced chip more expensive

😁 Hello, super humans! The most important number in tech this week is not a benchmark, it is a price list. TSMC quietly told its biggest customers that advanced wafers are going up, and since almost every chip you care about starts on one of those wafers, this is the kind of story that shows up in your cloud bill six months from now. Let’s dig in.

📰 Quick Signals

• 🧠 AI: SpaceXAI launched Grok 4.5, its first model “trained specifically for coding and agents,” pitched as Opus-class but cheaper at $2 per million input tokens, via its official announcement.
• 🤖 Robotics: Ex-Tesla scientist Rémi Cadene unveiled Paris-based UMA and its lightweight “Northstar” humanoid aimed at European factories, per The AI Insider.
• 💻 Programming: Kimi K2.7 Code became the first open-weight model selectable in GitHub Copilot’s model picker, announced in the GitHub changelog.
• ⚡ Electronics: An ETH Zurich team built a deterministic fuzzing framework for multi-hart RISC-V CPUs and found five previously unknown concurrency bugs in open-source cores, covered by Semiconductor Engineering.
• 📡 Telecom: Rocket Lab agreed to acquire Iridium and its 66-satellite low-Earth-orbit network in an $8 billion deal, per the Rocket Lab release.

🔍 The Big Story: TSMC just made every advanced chip more expensive

If you build anything that ends up running on silicon, from an AI service to a hobby board, the world’s most important foundry just moved the floor under your costs.

What happened: TSMC has notified major customers, including Nvidia, AMD, Apple and Qualcomm, that it will raise wafer prices by roughly 5 to 10 percent across its advanced nodes, which it defines as 7nm and everything more advanced. As Tom’s Hardware reports, that band accounts for about 74 percent of TSMC’s wafer revenue, so this is not a niche adjustment; it is a repricing of nearly the entire leading-edge market.

The details: The steepest move is at the top: TrendForce reports 3nm could climb up to 15 percent in the second half of 2026, with a further 5 to 10 percent penciled in for 2027. The reason is brutally simple supply and demand: Nvidia, AMD, Google and AWS are all racing onto 3nm for AI silicon, pushing monthly 3nm capacity from around 130,000 wafers at the start of the year to an estimated 160,000 to 175,000 in Q2, with utilization reportedly above 120 percent. When a foundry is that oversubscribed, price is the only release valve. The one countermove worth watching: Japan’s Rapidus is reportedly going after the same customers with aggressive 2nm pricing, targeting up to ¥3.5M per wafer, per TrendForce. TSMC reports Q2 earnings on July 16, which will show how much of this pricing power is already in the numbers.

🗞️ More News

🧠 AI

• OpenAI opened GPT-5.6 Sol, Terra and Luna to all users as its government-coordinated preview ended, and introduced an “ultra mode” that spins up subagents for hard tasks, per its launch post.
• Google DeepMind and partners opened a research funding call of up to $10M for multi-agent AI safety, with applications due August 8, detailed on the DeepMind blog.
• Anthropic redeployed Claude Fable 5 to general availability after the US lifted the export-control order that had pulled it offline, tracked on the Anthropic newsroom.
• DeepMind published its plan for defending internal systems against its own increasingly capable, imperfectly aligned agents, in a technical post.
• Google is reportedly delaying Gemini 3.5 Pro to July 17 for a full architectural rebuild, though it has not confirmed the date, according to Tech Times.
• Grok 4.5 ships trained alongside Cursor and priced at $2 in, $6 out per million tokens, undercutting frontier proprietary models on the SpaceXAI console.

🤖 Robotics

• Tesla is targeting a late-July start for Optimus Gen 3 production at Fremont, on a line converted from the old Model S and X, per Electrek.
• Boston Dynamics’ electric Atlas is advancing through Hyundai and Google DeepMind pilots on industrial tasks, covered by DigiTimes.
• UMA introduced “Real-Time Learning,” a system meant to let robots pick up tasks from demonstrations instead of manual programming, per The AI Insider.

💻 Programming

• Astral shipped its Rust-based Python type checker “ty” in beta, benchmarking 10 to 60 times faster than mypy and Pyright without caching, on the Astral blog.
• GitHub expanded Kimi K2.7 Code to Copilot Business and Enterprise, off by default until an admin enables the policy, per the GitHub changelog.
• Astral’s uv package manager is overtaking pip in CI for major projects and is prepping first-class support for Python 3.14 free-threaded interpreters, per the Astral 3.14 writeup.

⚡ Electronics

• TSMC and Amkor signed a 10-year U.S. advanced-packaging and test partnership centered on Arizona, tied to Amkor’s roughly $7B Peoria campus, in the Amkor announcement.
• Japan’s Rapidus is reportedly targeting up to ¥3.5M per wafer for 2nm to undercut TSMC as pricing power shifts to foundries, per TrendForce.
• The same ETH Zurich fuzzing work injected bugs automatically to benchmark existing RISC-V CPU fuzzers, a companion study on Semiconductor Engineering.

📡 Telecom

• Amazon Leo crossed roughly 390 satellites after a ULA Atlas V lofted 29 more, which the company says is enough to start initial broadband service later this year, per Amazon.
• Verizon says it now runs roughly 60,000 virtualized RAN sites, all from Samsung “because Ericsson is not ready,” as it pushes open, virtualized RAN toward 6G, per Light Reading.
• AST SpaceMobile is targeting an August launch for BlueBirds 11, 12 and 13 to grow its direct-to-device constellation, per Converge Digest.

👨‍💻 Code Corner

Today’s Big Story is a price change, so let’s quantify it. A wafer hike does not stay a wafer number; it lands on the cost of each working chip, scaled by how many good dies you cut from the wafer. This snippet estimates dies per 300mm wafer and turns a 10 percent wafer hike into a per-die number.

import math

def dies_per_wafer(wafer_mm=300, die_mm2=100.0):
    r = wafer_mm / 2
    # gross dies, corrected for wasted edge area
    return int((math.pi * r**2) / die_mm2
               - (math.pi * 2 * r) / math.sqrt(2 * die_mm2))

def cost_per_good_die(wafer_usd, yield_, wafer_mm=300, die_mm2=100.0):
    good = dies_per_wafer(wafer_mm, die_mm2) * yield_
    return wafer_usd / good

before = cost_per_good_die(20_000, 0.80)          # $20k wafer, 80% yield
after  = cost_per_good_die(20_000 * 1.10, 0.80)   # same wafer, +10%
print(f"${before:.2f} -> ${after:.2f} per good die "
      f"(+{(after / before - 1) * 100:.0f}%)")

The 10 percent wafer hike passes straight through: about $9.16 to $10.07 per good die on these assumptions. Nothing absorbs it unless yield improves.

🧰 Toolbox

• ty: Astral’s Rust-based Python type checker, now in beta and 10 to 60x faster than mypy.
• uv: the Rust package and project manager quietly replacing pip in CI pipelines.
• Cursor: the AI code editor SpaceXAI trained Grok 4.5 with, now a default surface for the new model.
• KiCad: the free, cross-platform EDA suite for designing the boards you would send off to be fabbed.
• INA219/INA226 breakout: a tiny I2C current and power monitor to measure what your gadgets actually draw.

🛠️ Build of the Week (rotating)

DIY USB power meter: wire a current-sensing breakout to a microcontroller and read live volts, amps and watts off any USB gadget on a small display.

• Difficulty: Beginner to Intermediate
• Parts: ESP32 or Pi Pico, INA219/INA226 breakout, small OLED, USB pass-through cable
• Why we like it: when silicon gets more expensive, knowing exactly how much power your build wastes is the cheapest optimization you own, and it makes today’s electronics posts tangible on your own bench.

📚 From the Blog

• How the Internet Stack Really Works: the layered journey from Enter key to rendered page, the same physical-and-protocol stack under today’s Rocket Lab and Amazon Leo satellite headlines.
• Resistance and Ohm’s Law: Controlling the Flow: resistance and Ohm’s law from the ground up, the math behind the power budget of every chip in the Big Story.
• What Electricity Actually Is: Charge, Current, and Voltage: charge, current and voltage from first principles, the physics running under every transistor TSMC just repriced.

😀 The Bot Says…

Moore’s Law promised the transistors would get cheaper every year. Nobody promised the wafers would. Beep boop: this bill’s on 3nm.

---

That’s all for today! If your roadmap assumes compute keeps getting cheaper, this is your reminder to pressure-test that assumption. Reply and tell us which node your project actually needs.