Charts

Hyperscaler capex is near $150B a quarter

The five biggest AI spenders, Microsoft, Amazon, Google, Meta and Oracle, are now laying out close to $150B of capital every quarter, almost double a year ago. This is the raw money behind the AI build, straight from earnings filings.

How much of the memory boom is AI

Korea's memory exports are HBM-heavy, the stacked memory bound for AI accelerators. Taiwan's are commodity DRAM and NAND. Both are the same product class in the same customs unit, so the gap between how fast each is growing isolates the AI premium riding on top of the ordinary memory cycle. When the two lines move together the boom is broad; when Korea pulls away, AI is doing the work.

What moved on the AI supply board this week

The seven-day change in every tracked constraint scarcity score. Positive bars tightened, negative bars eased. It is the fresh-signal companion to the buildout status board: not what is tight right now, but what changed this week. Straight from the world model.

The HBM premium did not collapse

A widely cited read of Korean customs data says the HBM premium over commodity DRAM has fallen to near parity. It has not. The blended customs number averages real HBM bound for TSMC packaging, around $200,000 a kilo, with commodity stacked memory re-exported through Hong Kong. Filter to the HBM-grade flow and the premium is still multiples above commodity DRAM. It compressed from a 2025 peak as commodity DRAM rallied fivefold, but real HBM never crashed.

DDR5 server memory is climbing

Spot prices for DDR5 server memory are up about 16% in 90 days and still accelerating. Rising memory prices are the clearest sign that the AI build is straining the supply that feeds it. TrendForce spot data tracks the move week by week.

Korea is exporting $26B of chips a month

South Korea, the world memory hub, is now shipping about $26B of semiconductors a month, more than double a year ago. Customs data is the hardest macro read on global AI and memory demand.

Phison rides the NAND shortage

Phison makes the controllers that turn raw NAND flash into usable SSDs, a quiet tollgate on AI storage. With NAND prices roughly doubling in six months, its monthly revenue has set records.

NAND flash is the tightest it has been

Tessara scores how scarce each input in the AI build is, from 0 to 100. NAND flash has climbed into critical territory as inference workloads pull storage demand and prices roughly doubled in six months.

HBM3e memory supply tightness

HBM3e is the high-bandwidth memory stacked beside every leading-edge AI accelerator. Tessara scores how scarce it is, from 0 to 100, updated daily.

DDR5 server memory supply tightness

DDR5 server DRAM is the commodity memory the AI build leans on. Rising tightness is one of the clearest signs the build is straining the supply that feeds it. Tessara scores it 0 to 100, daily.

Memory prices are at a record high

The memory super-cycle is not a forecast: it is already in the customs data. Korea's export unit prices for commodity DRAM and NAND are both at all-time highs, with DRAM up about 453% and NAND about 326% over the past year. The 2022 to 2023 memory glut is long gone; the AI build has bid the entire memory complex to record levels.

DRAM prices are breaking out

The export price of South Korean DRAM has more than quadrupled in a year and gone vertical, breaking to a new high after a long flat stretch. Korean customs unit prices are an early, hard read on the memory price cycle the whole AI build runs on.

NAND prices are surging

The export price of South Korean NAND flash is up more than 250% in a year, surging out of a long glut to a new high. After years of oversupply, flash has flipped to shortage as the AI build pulls memory of every kind.

Real HBM exports sell for six figures a kilo

The HBM bound for TSMC packaging leaves Korea at roughly $200,000 a kilo, multiples above commodity DRAM. It is the scarce memory stacked onto every AI accelerator, and the AI build is outbidding everyone for it. Korean customs unit prices on the Taiwan-bound flow track the real HBM price.

The value of HBM into Taiwan has more than doubled

The value of high bandwidth memory Korea ships to Taiwan has more than doubled in a year. It is the feedstock for TSMC CoWoS packaging, where HBM is stacked onto AI accelerators. Korean customs data tracks the pipeline that feeds advanced packaging.

HBM into Taiwan is getting pricier, not just bigger

The value of high bandwidth memory Korea ships to Taiwan has risen many times over since 2021, but the physical tonnage has grown far less. The difference is price: the AI build is not just buying more HBM for TSMC CoWoS packaging, it is paying multiples more per kilo for scarce supply. A supply squeeze shows up as price outrunning volume.

H100 rental prices are rising again

The price to rent an H100 GPU has climbed back to about $2.43 an hour, up roughly 10% in 90 days. After the 2025 glut, compute is tightening again. The Tessara compute price index tracks the neocloud rental market daily from provider rate cards.

MediaTek is building an AI-silicon business

MediaTek, best known for the chips in a huge share of the world's phones, is now designing custom AI accelerators for hyperscalers. Its monthly revenue tracks how fast that data-center business is scaling on top of its core mobile line.

Which constraint is binding right now

Tessara scores how tight compute and memory are, each on a 0 to 100 scale. When one score outruns the other, it names the binding constraint for the AI build this week.

Auras keeps dense AI racks from overheating

As AI racks get too dense for air, more of the rack moves to liquid cooling, and that is the Auras lane: cold plates, manifolds, and rack-level thermal modules. Its monthly revenue tracks the shift to liquid.

Liquid cooling supply tightness

As AI racks get too dense for air, the cooling layer tightens. Tessara scores how scarce direct liquid cooling is, from 0 to 100, updated daily.

Where AI power is cheap, and where it is brutal

Spot electricity ranges several-fold across the grids where AI compute wants to live. That spread, not the average price, is what decides where the next data centers actually get built.

Power is the binding constraint

GE Vernova now sits on a 100 gigawatt gas-turbine backlog. When data centers cannot get grid power fast enough, they order their own generation. The backlog is a read on how hard power has become to secure.

Power transformer prices stepped sharply higher

US producer prices for power transformers are up about 4% in a year and stepped sharply higher in late 2025. Transformers are the long-lead chokepoint for connecting new data centers to the grid, so their pricing power reads how tight grid equipment has become.

US power grid spending keeps climbing

US construction spending on electric power infrastructure has climbed past $168B a year, up about 7% from a year ago and rising every month. The grid buildout is the hard-dollar backbone of the AI power crunch.

Imports of AI power gear keep climbing

US customs imports of gas turbines and HVDC converters, the equipment that gets AI data centers onto the grid, keep climbing. Gas turbine imports have gone from near nothing in early 2024 to tens of millions of dollars a month, with HVDC rising alongside. The hardest customs read on the power buildout, next to the server hardware already pouring in.

EUV demand just hit a new high

ASML sold about EUR 4.1B of EUV lithography systems last quarter, a record. Every advanced chip in the world is made on an EUV machine, and ASML is the only company that builds them. This is the cleanest read on whether leading edge supply is expanding.

TSMC revenue is up about 18% in a year

TSMC, the company that makes nearly every advanced AI chip, is now pulling about $13B of revenue a month, up about 18% from a year ago. Its monthly sales are the cleanest single read on AI chip demand.

Topoint drills every AI server board

Topoint makes the tungsten-carbide drill bits that punch the holes in AI-server circuit boards. As boards move to harder, higher-layer materials, demand for its bits keeps rising, and it is one of only two high-end suppliers in the world.

Nan Ya PCB makes the board GPUs sit on

Nan Ya PCB makes FC-BGA advanced substrates, the dense little board a GPU sits directly on top of. Substrates are sold out across the industry, and its monthly revenue tracks the shift from commodity PCB to advanced substrate.

CoWoS packaging stays tight

CoWoS is the advanced packaging that fuses a GPU to its memory, and it has been a binding chokepoint of the AI build. Tessara scores how scarce it is, from 0 to 100, updated daily.

Advanced substrate supply tightness

FC-BGA advanced substrate is the dense board every AI accelerator sits on, and it has been sold out across the industry. Tessara scores how scarce it is, from 0 to 100, updated daily.

Chip prices are rising again

US producer prices for semiconductors, which historically only fall, are now up about 26% in a year and breaking out. Rising chip prices are a clear sign that supply has tightened against AI demand.

Advanced-packaging tool demand hit a new high

Disco makes the precision grinders and dicers that thin and cut advanced AI chips. Its quarterly shipments just broke their prior peak, up about 28% in a year. Disco sits upstream of CoWoS packaging, so its order book is an early read on AI accelerator supply.

WIN Semiconductors fabs the lasers for optical AI links

WIN is a III-V compound-semiconductor foundry that fabricates the indium-phosphide lasers carrying data between racks as light. As co-packaged optics moves inside the rack, its monthly revenue tracks the early ramp.

US chip imports from Taiwan have doubled

The value of chips the US imports from Taiwan has more than doubled in a year, to about $1.7B a month. Customs filings are the hardest evidence of US AI infrastructure pulling demand through TSMC.

Taiwan AI-server exports keep climbing

Taiwan now exports more than $8B of AI servers a month, up close to 150% from a year ago and many times the level of two years ago. These are the assembled machines that fill AI data centers, straight from Taiwan customs.

Wiwynn builds the racks AI runs on

Wiwynn is the ODM that integrates compute, cooling, power, and networking into the finished racks hyperscalers deploy. Its monthly revenue is a direct read on how fast AI server systems are shipping.

Seven Taiwan names inside the AI supply squeeze

A live screener of seven Taiwan suppliers that each sit on a tight rung of the AI build, from NAND controllers to liquid cooling to the substrate a GPU sits on. For every name it shows our supply-tightness score, market cap, forward P/E, and the latest monthly revenue, sorted cheapest first. The numbers update on their own as the data moves.

The AI build, in hard customs dollars

America is importing more than $23B of server hardware a month, up close to 190% in a year. Customs filings are the hardest evidence that the buildout is real spending, not just press releases.

Two hyperscalers are outspending their cash

Capex has overtaken operating cash flow at Amazon and Oracle, so both now fund the AI buildout with debt and reserves. Meta, Microsoft and Google still cover capex from cash. The ratio is the cleanest read on who can keep spending at this pace.

Global chip sales are surging

Worldwide semiconductor sales have climbed about 62% in a year to roughly $89B a month. SIA monthly sales are the broadest read on chip demand, the top of the funnel for everything the AI build consumes.

The AI buildout is rotating downstream

Across the companies we track by ocean shipment, the wave has moved from building chip factories to assembling the servers, cooling them, and wiring up the power. You can see it in the freight before it shows up in revenue.

AI Buildout Status

The live supply board for the AI buildout: which inputs are tight or tighter right now, the single constraint binding hardest this week, and the share of the tracked supply lane under stress. It reads straight from the world model and updates as the scores move.