Chips, fabs, batteries and the physical machines the digital world is quietly built on.
At SEMICON Taiwan this week the industry conceded its fastest accelerators are starving for bandwidth. The fix is to route light through the package, and it hands a new chokepoint to a very short list of firms clustered around one island.
At two chip conferences this month, Samsung and SK hynix showed how they plan to keep stacking HBM for AI accelerators — and revealed how little physical room is left before packaging, not lithography, becomes the chokepoint.
Six years, 460 megawatts, one campus in West Virginia. Price the contract per kilowatt-hour and the electricity is the cheapest thing in it — which is exactly why it decides the whole deal.
Unitree's 500% debut and Beijing's robot games are the spectacle. The bill of materials is the story — 40 to 70 actuators per machine, and a supply chain that runs, link by link, back to one country.
Building-integrated solar was a bet that a beautiful tile could beat a commodity panel on cost per watt. The panel fell ninety percent. The tile needed a specialist to install it. The invoice decided the rest.
Its chip can run one kind of AI model and nothing else. That is the whole product — and, if the architecture that rules the field ever shifts, the whole risk.
Behind-the-meter gas isn't cheaper power. It's faster power — and a bill in Washington would let it skip the rules entirely. The cost still shows up. Just not on the data center's meter.
Trump's new proclamation hits finished drones in 21 days. The motors, magnets, batteries and camera modules that make one fly are Chinese too — and those got 180 days, because everyone knows they can't be re-sourced by then.
A binding $6 billion joint venture in Kumamoto ends Sony's in-house era and routes the world's image sensors through the same advanced-manufacturing chokepoint as everything else. The reason is a wall Sony hit in the layer you never see.
Tesla and SpaceX have committed $16.8 billion to Terafab, a Texas plant meant to make logic, memory and packaging under one roof. Every one of those is a separate decade of craft — and the company's own IPO filing calls the whole thing a framework with no binding commitments.
A $15 billion, one-gigawatt AI hub is going up next to a wildlife sanctuary in Visakhapatnam, a city that gets 410 million litres of water a day against a 480 million-litre need. The promise is air cooling and its own power. The number to watch is the gap.
HP, Asus and Acer are reportedly putting CXMT's Chinese DRAM into budget laptops sold outside the US. They aren't doing it to save money. They're doing it because the shelf is empty.
Trump's Ratepayer Protection Pledge now covers eighty percent of America's electricity and commits no utility to anything. The number that decides your bill isn't in it — it's the 4.4 billion dollars already spread across ratepayers, and the phrase "directly attributable."
High-bandwidth memory has entered high-volume production, and the three companies that make the world's DRAM are quietly walking away from the market that used to define them. This is not a shortage. It is a reallocation, and reallocations do not reverse on their own.
This week's earnings put hyperscaler spending near a hundred and forty billion dollars a year. The 'clean nuclear' answer is a fraction of one percent of that, at three times the price of wind — and it mostly delivers after 2032. The grid runs on gas until then.
Qualcomm is raising chip prices by double digits on September 1. The cause sits two chokepoints upstream — a fab that would rather print AI accelerators, and three companies that would rather sell their memory to data centers.
Frozen v2 is reportedly six to ten times more efficient per watt than a Google TPU, because it hardwires the model's shape into the hardware. In a world where the grid, not the fab, is the ceiling, efficiency is the only lever left. It is also a treadmill.
Iris is designed on Broadcom's platform, fabricated by TSMC, and fed by memory Meta buys from Samsung. Building your own accelerator no longer means leaving the queue — it means choosing which queue you stand in.
A one-year moratorium on anything drawing 50 megawatts or more is really a bill coming due: the state wants the load to pay for the grid it strains, instead of the household on the same feeder. Whether it works comes down to one number nobody has set yet.
The NT$1.27 trillion headline hides the two forces even the world's most important chipmaker cannot control — a strengthening home currency and a packaging line that is already sold out — and one of them is worth three times the entire revenue beat.
Everyone counted the fabs in Thursday's announcement. The line worth reading is the financing for a wafer plant in Sherman, Texas — because a fab without blank wafers is a very expensive building, and nearly every one of those discs on Earth comes from a handful of companies, none of them American.
The honest number in Google's 2026 environmental report isn't the one it leads with. It's the gap between the electricity Google bought on paper and the electricity the grid actually burned to keep its AI running.
Meituan's LongCat-2.0 is the first frontier-scale model trained end to end without Western silicon. The achievement is real. What it hides is where the chokepoint actually moved.
onsemi is paying $7 billion in its own stock for Synaptics. The deal is a bet that intelligence is about to leave the cleanroom and ride into the car, the robot and the factory sensor — and that whoever owns the dull silicon around it owns the volume.
The largest corporate investment pledge in Korean history is mostly chip fabs and AI data centers. The country they sit in is an electrical island, importing almost all its fuel, with more than half its planned grid upgrades already delayed. The cheap part was the won.
Record revenue, an 85% margin, and $22 billion from customers paying cash up front. Underneath the AI memory boom is a manufacturing chokepoint deciding who gets DRAM — and who waits.
Tesla and NatPower will build 25 gigawatt-hours in Italy and Britain. The number that decides whether it matters isn't the capacity — it's the cost, and what the battery does between charges.
HBM4E is 16 gigabits a second per pin. It is also more than 20 percent more power-efficient and 17 percent cooler, and inside an AI data centre that second pair of numbers is the one that shows up on the electricity bill.
18A-P is a real engineering step. But a leading-edge node without an outside anchor customer is a stranded asset — and the name on Intel's wish list is Apple.
The Commerce Department's CHIPS office is funding SandboxAQ to invent PFAS-free coolants, rare-earth-free magnets and domestic catalysts in software. Designing the molecule was never the slow part. Qualifying it into a fab is.
Schneider Electric and Foxconn want to stamp out power-and-cooling modules the way Foxconn stamps out servers. It compresses the build. It does not move the number that actually gates AI: megawatts you can get to the site.
Blocked from selling its GPUs, Nvidia is pitching Chinese clouds an 88-core CPU instead. It slips through a line drawn around the wrong half of the machine — and the catch is what the chip cannot do without the other half.
Beijing is reportedly drafting a five-year, state-directed data-center plan built around domestic silicon and a near-total squeeze on Nvidia. The money is the easy part. The constraint sits one layer below the logic, in a component China still cannot make enough of.
Most of the US data-center capacity promised for 2026 and 2027 has not broken ground. Not for want of chips or money, but because the grid runs on a slower, more physical clock than capital does.
At Computex 2026, the x86 establishment answered an AI-PC era being redrawn by Arm and Nvidia. The pitch was confident. The manufacturing reality, as always, was the harder story.
Nvidia and MediaTek will show an Arm laptop chip in Taipei this week and call it a break with the past. Follow where it is actually designed, fabricated and packaged, and it looks less like a new era than a deeper turn of the same dependency.
A queue four years deep, a battery boom measured in hundreds of gigawatt-hours, and one unsettled question: who pays the bill.
Sodium, iron and air are not exciting. They are abundant, cheap and durable. That, not energy density, is why they are taking the grid.
When TSMC says "1.6 nanometres," it is no longer describing a transistor. It is describing a system — power fed from the back, dies stitched together on a wafer, heat that has nowhere to go. Read the node alongside the machine, or you have not read it at all.
Everyone watches the transistor count. The AI compute supply chain is actually rationed somewhere quieter — on a slab of silicon the size of a dinner plate, in a process called CoWoS, where a single company decides who gets to build the future and who waits.
