First the SaaS-pocalypse. Now the Chip-pocalypse. Same panic, Same bounce?
Back in March, we were all told Software was dead, now the very thing that was going to make it obsolete is under the same scrutiny. The panic is familiar, the moves are similar, and we think the inevitable bounce will be the same once the dust settles.
The 30 stocks in the PHLX Semiconductor Index (NYSE: SOXX) have lost a combined US$6.5 trillion from their 2026 peaks. Down 35%. NVIDIA accounts for US$2 trillion of that fall on its own. US$1.78 trillion has gone since the US midterms on 3 November. Twenty-six of the 30 stocks are lower since election day.
The market's verdict is swift, simple and almost unanimous. Democrats won. Democrats hate data centres. The AI build-out is finished. No multiple is safe.
SOX has lost US$6.5 trillion from its 2026 peaks. Forward PE is the lowest since ChatGPT launched. Data centres aren't being cancelled. They're being moved. Or hyperscalers will deal direct. And just like "Sass-pocalypse" the chip stocks will be a buy.
SOX top 10 have lost US$5.6 trillion from their own 2026 peaks. NVIDIA is 36% of that.
How much has actually been cancelled?
Not much. We tracked every data-centre and chip cancellation we could find between 3 and 17 November. Only two carry a dollar figure.
Confirmed: Lighthouse Sustainability pulled a US$10 billion campus in Delhi Township, Michigan. Reported, not confirmed: the FT says a US$7 billion, five-year lease of Nvidia chips between Oracle and Tencent was scrapped over export controls. A lease, not an order cancelled with Nvidia.
That's US$17 billion against US$1.78 trillion of market value gone since the midterms. More than a hundred to one. Measured from the peaks, it's about 382 to 1.
And not one confirmed GPU, HBM, DRAM or NAND order cancellation has surfaced. Not one (yet).
Data centres don't die, they move
This is the bit the market is missing. Look at what actually happened to the "cancelled" projects.
Telus lost its partner on two Vancouver AI sites, planned for around 63,000 Nvidia chips. The developer, Westbank, says it's going ahead with someone else. In the same fortnight Google committed US$17 billion in Wyoming and US$15 billion in Malaysia. That's US$32 billion of new projects, almost double every valued cancellation combined. Meta's president has already said every hyperscaler is asking whether it needs to build outside the US.
A blocked site in Ohio doesn't destroy a GPU order. It sends the GPU to another state, another campus or another country.
And where the hyperscalers do stay, I reckon they'll end up dealing directly with counties and states. Bring your own power, pay for the grid upgrade, sign a community agreement, get your approval. Dublin has restricted new grid connections since around 2022 and now wants new sites to bring their own power. Singapore paused around 2019 and reopened roughly three years later with selective approvals. Pause, then rules, then slower and dearer growth. That's the pattern. Not extinction.
Then there's the money already locked in. Microsoft's (NYSE: MSFT) latest 10-Q shows US$364.6 billion of leases signed but not yet started, mostly data centres. Meta reports around US$355.7 billion. Amazon has kept its 2026 capex guidance at about US$220 billion, Alphabet at US$195 to 205 billion. These are contracts, not press releases.
And the existing market is tight. CBRE's mid-2026 survey had 7,481 MW under construction across eight primary North American markets, 80.4% pre-leased, with vacancy at 1.4%. When vacancy is that low, every delayed building makes the powered ones worth more, not less.
Politicians' election promises versus reality is low
Partly the bears are right, and we're not dismissing it.
Voters in Trenton and Piqua, Ohio, banned data centres over 25 MW on election day. JPMorgan estimates 37.5 GW of planned 2028 US capacity could face pauses. Nvidia's Vera Rubin ramp has reportedly slipped to 1H27 on an HBM4 bottleneck.
But exposure isn't cancellation. When Jefferies, the broker that called the midterms the "worst-case near-term scenario", actually sharpened its pencil, it cut its forecast for incremental US capacity through 2030 from 122.4 GW to 112.8 GW. That's a haircut of about 8%.
So we ran our own worst case. Every one of JPMorgan's 37.5 GW paused, plus every blocked, delayed and undisclosed project in our tracker, at US$35 million per MW all-in, chips included. That's about US$1.5 trillion of capex, spread over several years, around half of it chips. It's still smaller than the US$1.78 trillion the SOX has lost since the midterms. And remember it assumes none of that capacity gets moved or approved later, which is exactly what we think will happen.
The bigger risk for us is China. About one in six SOX revenue dollars (16.7%, or US$140.7 billion) is billed to China including Hong Kong, the group most exposed to US export bans. A new BIS rule adds licence requirements for Malaysia, Singapore, Thailand, Indonesia and Brazil. That one we'd watch more closely than any county vote. But it is still a fraction of total demand.
1 in 6 SOX revenue dollars comes from China, but only $25b falls under export controls
Of the 16.7% of SOX revenue billed to China, only about 45% of the equipment makers’ slice is classified as national-security restricted (advanced-node tools, Entity List end-users and similar). The rest of that China exposure is ordinary commercial demand the export rules do not cover.
SOX forward PE is the lowest since ChatGPT launched
Cheap. On our numbers the SOX is on 13.4x next-12-month earnings, the cheapest since ChatGPT launched. 18 of the 30 SOX stocks trade on under 20x forward earnings, nine are under 15x, and 13 are down more than 50% from their 2026 peaks. NVIDIA is on 9.90x.
SOX at cheapest forward PE since ChatGPT launched.
But not everything is cheap for the right reason. The 30 stocks are still up about 12% combined this year, so this is a give-back of mid-year gains, not a wipe-out. And Micron on 3.41x after a 152% run this year is a reminder that a tiny multiple at peak memory earnings has bitten people before.
Year-to-date return vs fall from 2026 peak, ten lowest forward P/Es.
How we're playing it
NVIDIA (NASDAQ: NVDA): 9.90x forward, down 16.3% this year and 34.8% from its October peak. It's about 69% of the SOX fall since the midterms, so most of the fear sits in one stock. The leadership change is a real unknown. But a sub-10x multiple on the AI platform leader, with no confirmed GPU cancellations, prices in a collapse the order book isn't showing. Catalyst: results this week.
Taiwan Semiconductor (NYSE: TSM): 16.2x, 23.2% off its peak. Whatever state or country the data centre ends up in, the leading-edge chips still come out of a TSMC fab, so the "they'll just move" thesis doesn't hurt it. Risk: Taiwan itself.
Micron (NASDAQ: MU): 3.41x, the cheapest in the index. For stronger nerves only. Cycle risk, and its US$4.5 billion Fab 11E in Taiwan is paused over a union pay dispute. Size it accordingly.
What would change our mind?
Worse: Nvidia, TSMC or a hyperscaler confirming lower orders, a capex guidance cut, or a memory maker reporting cancelled HBM or DRAM orders.
Better: Nvidia's results holding up under interim leadership, OpenAI's Camellia project approved in Georgia, Micron's Fab 11E restarting, and the first big state-level "bring your own power" deal getting signed.
The blue wave will make the next tranche of US data centres slower, dearer and more political. It won't stop the chips shipping. The news cycle will move on long before the policy cycle does.
The cancellations are real. They're just more than a hundred times smaller than the sell-off.
5 topics
5 stocks mentioned