Explained
AI Tokens vs GPU Hours: Same Chip, Different Bills
Tokens are how a cloud bills the model. GPU hours are how it bills the card. They only match at a conversion ratio — billed tokens per live GPU-hour.
Sean
Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Not investment advice.
Wire
NVIDIA pays $1.1bn for US test as Amkor deals hit $3bn
NVIDIA said at GTC Berlin that it is booking substrate, packaging, test and co-packaged-optics capacity with six suppliers. The named cash is a $1.1 billion KYEC advance and Amkor agreements that now total $3 billion.
Wire
NextEra's 16.5 GW of data center power is funded, not yet sold
NextEra says three federal gas hubs in Texas and Pennsylvania, about 16.5 GW, have $5.7B of first funding from Japan and Korea. NextEra earns fees without capital. The data center buyers who are meant to pay for the power have not signed yet.
AI Data Centers
What Does TeraWulf Do? Bitcoin Miner to AI Data Center Landlord
TeraWulf is a former Bitcoin miner that now leases powered AI data center halls: 839 MW signed with Core42, Fluidstack (backed by Google) and Anthropic, 102 MW earning rent, and a 1 GW Kentucky campus with power contracted but no tenant yet.
Wire
Nscale talks a second $45B, 460 MW Anthropic data center deal
A second 460 MW Monarch building would make Anthropic about 60% of Nscale's book. The first $44.6 billion contract still has no binding financing, and late tranches can be cancelled.
Catalyst
Nvidia books 2.5x substrates and 300% more interposer capacity
Nvidia is paying suppliers in advance for 2.5 times more substrate capacity and more than 300% more interposer capacity by 2028, which shows the steps after the wafer are now what Nvidia spends its own cash to unblock.
Explained
Who Are the Hyperscalers? The Companies Paying for the AI Build
The hyperscalers are the few companies that build and run computing at the largest scale. Four US names spent about $166 billion on capacity in one quarter. Chips, landlords and power get paid first; the buyers wait on power-ready halls.
China
10,000 of 30,000 Cards: Who Gets Paid While China's First Ascend 950DT Pool Waits on a Proof Run
The first 10,000-card pool of Huawei's new Ascend 950DT chips is due by the end of October in a hall built for 30,000. The building, the power and the cooling are done. What nobody has shown yet is a weeks-long training run on it.
Analysis
What Does Lam Research's Price Assume About the China-Licence Stuck Step?
Lam Research closed at $263.45, about $329.7 billion. It already shipped $4.250 billion of etch tools. What does the price assume about getting paid on the etch while two China NAND fabs wait on a US licence?
Analysis
Wiwynn's Sales Doubled but Cash Went Out. What Does the Price Assume?
Wiwynn's September revenue doubled, yet its operations used NT$61.6 billion of cash in the first half as inventory and memory bought for customers piled up. What the price assumes, who gets paid first, three scenarios with the arithmetic shown.
Wire
Samsung moves Xian to 286-layer NAND as SK Hynix eyes 240 at Dalian
Samsung targets 40,000 wafers a month of 286-layer NAND at Xian by December 2027 and SK Hynix wants 240 layers at Dalian. How far the US has licensed the tools is undisclosed.
Hyperscalers
How Do Hyperscalers Buy Power? The PPA and Own-Plant Model
A hyperscaler buys megawatts the way it buys chips: it signs first, pays before the plant is live, and only then can a hall earn.