Hardware Outlook · Canada · October 31, 2026
The investor who shorted the 2008 housing market now holds put options against Nvidia, Palantir, and Micron, the company that makes a large share of the world’s RAM. His bet and your RAM bill both run through Micron.
“feels like the last months of the 1999-2000 bubble”
— Michael Burry
In late September, Michael Burry covered his outright shorts in Nvidia, Micron, and Palantir and replaced them with put options in all three, telling subscribers he expects a 1987-style market crash. The Nvidia and Palantir positions extend a bet he first disclosed a year ago, when his fund Scion Asset Management reported put options with notional values of roughly $187 million on Nvidia and $912 million on Palantir. Micron is the new position, and the one that matters most to gamers. Micron fabricates a large share of the DRAM in your PC. Its stock has nearly tripled this year on memory demand from AI datacenters, and it is the same demand that has roughly tripled RAM prices in under a year. A 32GB Crucial kit that sold for pocket change in early 2025 now runs $369 on Amazon, per SlashGear’s October build comparison, and the RTX 5090 has climbed past $7,000 against a $1,999 MSRP.
Burry’s target is the AI buildout itself. He is predicting that the spending spree behind the shortage ends, and that the companies funding it pay for it in public markets. But the mechanism he is betting against, hyperscaler capital expenditure flowing into chips and memory, is the exact column holding up every component price you have been watching climb. If the column breaks, the shortage does not end politely in 2028 as the memory makers forecast. It ends whenever it breaks, and memory markets have a long record of ending their booms violently.
The questions worth settling are what a pop would do to each part of a gaming build, in what order and after what delay, and how deep the last three comparable crashes went. Then there is what you should do with a parts list and a budget while the bet is unresolved, because Burry’s puts expire between June and December 2027, which is also the window where every forecast for this market bends.
Burry has been warning that this rally “feels like the last months of the 1999-2000 bubble” since May, when he told Substack readers to “reject greed” and cut parabolic positions almost entirely. The market did not listen, and the Philadelphia Semiconductor Index rose nearly 70 percent between late March and mid-May alone. His own position has escalated in stages since November 2025.
| Date | Move | Size / instrument | Thesis element |
|---|---|---|---|
| Nov 2025 | 13F discloses puts | ~$187M notional on Nvidia, ~$912M on Palantir (CNBC) | AI valuations are dotcom-scale; financing is circular |
| May 2026 | Substack warnings | n/a | “Last months of the 1999-2000 bubble”; raise cash (Investopedia) |
| Jul 2026 | Shorts Micron | outright short (Seeking Alpha) | Memory boom-bust cycle is unchanged; AI is “not nearly different enough”; Chinese supply and new fabs will catch demand (Motley Fool) |
| Sep 2026 | Covers shorts, buys puts on NVDA, MU, PLTR | puts expiring Jun-Dec 2027 (Notebookcheck) | Predicts a 1987-style crash “sooner than later”; GPUs depreciate faster than the books imply |
Burry’s record since 2008 is full of early calls that were wrong on timing, and he runs a paid newsletter that benefits from a readership refreshing for the next warning. He also has company: Palantir CEO Alex Karp dismissed him as “bats crazy,” and the bull case is real. The three targeted companies, Nvidia, Micron, and Palantir, sell into genuine recurring demand, and their biggest customers, Alphabet, Microsoft, Amazon, Meta, and Oracle, are collectively profitable. None of that changes the useful part for a gamer. A gamer can plan around the possibility without picking a side. You need to know what his being right would do to a parts list, and what signals would tell you it is happening before the retail prices move.
The fabs are running at full tilt. It is a demand problem: AI datacenters are bidding the wafer output away from consumer memory, and the size of that bid is set by one number, hyperscaler capital expenditure. In 2026, Amazon, Alphabet, Meta, and Microsoft plan combined capex of $720 billion to $745 billion, and broader estimates that include other builders reach up to $828 billion. That is the money paying for the HBM and server DRAM that Samsung, SK hynix, and Micron prioritize over the DDR5 and GDDR7 that go into gaming PCs.
The transmission into your prices has been brutal and fast, and it is already decelerating.
The retail numbers translate directly. Tom’s Hardware reported RAM prices tripling within a year, a figure our drought coverage cited in February, and the increases have continued since: SlashGear’s October price comparison found that same $369 32GB Crucial kit in a budget build where memory used to be an afterthought. GPUs carry the same pressure through their GDDR7: the RTX 5090 streets around $7,000 against a $1,999 MSRP, and the 16GB RTX 5080 above $2,000 against $999 (Digital Foundry, Wccftech). Both premiums are scarcity prices, and scarcity prices are the first to collapse once the demand behind them fades.
TrendForce itself has started naming the risk. In its July outlook it noted that some cloud providers are reporting negative free cash flow despite record spending, and that whether memory costs “ultimately affects CSP investment plans and memory procurement remains a key variable to watch.” Leaker Moore’s Law Is Dead puts it more colorfully, describing an “AI debt bomb” in which companies like CoreWeave use GPU infrastructure as collateral to borrow money for more GPUs, and argues Nvidia is diverting consumer silicon into AI and workstation cards to sell as much as possible before the music stops (Notebookcheck). The collateral loop matters because it is where a stock-market crash becomes a hardware-market crash.
A crash in AI stocks does not make RAM cheap the next morning. Memory is a contract market buffered by long-term agreements, supplier inventory, and utilization decisions. But it also has a spot market that front-runs contracts, and history shows retail following spot within weeks. The chain runs through five steps.
One detail specific to this cycle: GPUs have a second, weirder channel. Moore’s Law Is Dead reports GB202 dies are currently flowing from GeForce into RTX Pro workstation and AI cards. If AI demand dies, that allocation reverses. The silicon that was supposed to be an RTX 5090 comes back to the gaming channel, which is the one scenario where a demand crash simultaneously increases consumer GPU supply.
Three historical episodes bracket the plausible size of a fall. They differ in cause, and the differences matter, but the magnitudes cluster in a narrow and brutal band.
The 2019 DRAM downturn is the closest analog, because it was datacenter-led. Server buyers who had overbought through 2018 stopped, inventories hit records on both sides of the table, and prices fell more than 40 percent over the year with individual quarters dropping 30 to 35 percent. The 2022 GPU crash is the closest analog for the used-market flood: when Ethereum mining economics died, eBay absorbed the liquidation and prices fell 57 percent in roughly six months, with high-end flagships falling hardest. The dotcom comparison Burry himself reaches for adds one more data point, at the stock level: Cisco was a profitable company in 2000, and its shares and Intel’s took more than a quarter century to recover, because when the customers stop buying, the supplier’s fundamentals stop mattering.
Two differences between those crashes and a 2027 pop cut in opposite directions. Against the bear case: today’s market is physically short, with TrendForce estimating the DRAM sufficiency ratio at −1 to −2 percent in 2026, so a demand collapse would hit supply that has no slack to dump, and memory makers have learned to cut utilization fast. For the bear case: consumer demand has been crushed by the prices themselves (PC shipments fell over 20 percent in Q3 on RAM costs, per The Verge), so there is no consumer floor to catch falling contract prices, and the HBM-heavy capacity mix is exactly the capacity an AI downturn idles first. The net read from history is that a fast fall of 40 to 60 percent off peak for RAM within three to five quarters of a genuine capex collapse is the precedent-consistent range, with GPU street premiums and used cards falling faster and further than new-card MSRPs.
The forecast split among people paid to know is wide. Micron’s CEO expects supply-demand to be “much tighter in fiscal 2027 and 2028 than they were in 2026,” and SK hynix has warned of undersupply past 2030. On the other side, the Acer CEO calls a shortage to 2030 “impossible” and expects relief around mid-2027, and a former president of Samsung’s semiconductor division forecasts a Chinese-driven supply surge in the second half of 2027 or early 2028, explicitly adding “a possibility of reduced investment” in AI. Burry’s scenario is the extreme version of that last view, with reduced investment arriving before the supply surge.
Hyperscalers keep spending but the growth rate flattens as negative-free-cash-flow providers pull back. The QoQ increases keep decelerating along the curve in the first chart, contract prices plateau in 2027, and the new capacity TrendForce expects from the second half of 2027 grinds prices down toward 2028. RAM ends the window roughly 25 to 35 percent below today’s peak. GPU premiums fade slowly as the incoming 24GB RTX 5080 normalizes supply, which is the path TrendForce’s own decelerating numbers currently point toward.
AI equities crack, the collateralized financing loop breaks, and 2027 capex guidance gets cut. Following the 2019 and 2022 precedents, spot DDR5 prices lead the collapse within weeks, retail follows within a quarter or two, and the 32GB kit lands somewhere in the $150 to $200 band by late 2027, a 45 to 60 percent fall. Used GPUs fall faster, the $7,000 RTX 5090 premium evaporates first of all, and diverted GB202 dies flow back toward GeForce. But the crash also freezes fab expansion, so the cheap window may only stay open a year or two before supply cuts re-tighten the market. In this scenario, patience pays, but only patience that was positioned before the signals turned public.
Capex holds at the $720 to $828 billion range or climbs, the SK hynix and Citrini view wins, and memory stays tight past 2027, possibly to 2030. Prices hold or drift higher into 2027, and the buyer who waited for a crash pays the full shortage price when they finally buy. This is the scenario our earlier drought coverage assumed, and it is why the default advice has been to buy at a fair price when you see one rather than wait.
The exposure in every scenario concentrates in memory, with RAM and SSDs hit hardest and the rest of a build largely untouched.
| Part | AI-demand exposure | First channel hit in a pop | Expected move in Burry scenario |
|---|---|---|---|
| RAM (DDR5) | Highest | Spot market, within weeks | −40 to −60% from peak; the most exposed part of any build |
| SSD (NAND) | High | Contract renegotiation; TrendForce already sees NAND loosening in 2H27 | −30 to −50%; falls with or without a crash, faster with one |
| New GPUs | Medium | Street premium evaporates before MSRP moves | MSRP roughly sticky; the 2.0 to 3.5x premiums over MSRP are what collapses, plus die reallocation back to GeForce |
| Used GPUs | High | Liquidation flood (2022 precedent: −57%) | −30 to −50% in months; flagships and ex-AI-workstation cards fall hardest |
| CPU / platform | Low | n/a; Nova Lake and Zen 6 arrive 2027 regardless | Little direct effect; Black Friday discounting proceeds normally |
| Case, PSU, cooling | None | n/a | No effect; safe to buy on promotion |
Retail prices are the last thing to move in this chain, and these signals move ahead of them.
| Signal | Reading now (Oct 2026) | What would confirm the pop |
|---|---|---|
| Hyperscaler capex guidance | Big 4 at $720-745B for 2026, still being raised | Any two of Amazon, Microsoft, Alphabet, Meta guiding 2027 capex flat or down on earnings calls. This is the load-bearing number; everything downstream follows it |
| DDR5 spot vs contract price | Spot above contract, both rising | Spot breaking below contract and falling for 4+ consecutive weeks. Spot led the 2019 and 2022 declines by weeks to a quarter |
| H100 rental rates | Down roughly two-thirds from late-2024 highs; one TechInsights estimate cited on investor forums put a further 30 to 40% spot decline since spring | Rates falling through $2/hr with occupancy dropping. Rental rates are the real-time demand gauge for the GPUs backing the memory orders |
| Memory-maker capex and utilization | All three majors expanding for AI demand | Samsung, SK hynix or Micron announcing utilization cuts or delayed fab equipment installs, the 2019 and 2023 playbook |
| Neocloud financing | CoreWeave-style collateralized borrowing continues | A major GPU-debt borrower missing a covenant, restructuring, or failing to roll debt. This is the “AI debt bomb” detonation MLID describes |
| Burry’s own positions | Puts on NVDA, MU, PLTR, expiring Jun-Dec 2027 | Quarterly 13F filings show him extending, rolling, or exiting. His covering the 2026 shorts to buy puts was a conviction increase; the reverse would be a conviction cut |
Buying a non-memory part now costs you nothing relative to waiting if the bubble pops. RAM bought at peak carries a real cost. A pop means you overpaid by the crash margin, which precedent says is 40 to 60 percent. Waiting is not free either, because if the bubble holds, you overpaid by the continued rise, which TrendForce’s own decelerating curve suggests is now in the 10 to 20 percent per quarter range and slowing. The downside of buying too early is larger than the downside of waiting too long, but only for memory parts, and only if you have the discipline to act on signals rather than hope.
In the shortage scenario, prices fall when the new fabs land, and TrendForce puts meaningful output from them in 2028. The Burry scenario delivers the same relief through demand dying instead, with his puts written to resolve between June and December 2027. They are two opposite theses about the market, aimed at the same window. The difference is the path, either a grind you can plan around or a crash that pays whoever kept dry powder and watched the spot market.
What separates the two paths on the way in is a single number that updates in public four times a year: the capex guidance of Amazon, Microsoft, Alphabet, and Meta. Until one of those companies says it is spending less, the shortage is real and buying at a fair price remains the right move, which is the same advice we gave in our drought coverage. The moment two of them say it, the fan chart’s red band comes into play, and the playbook flips from “buy fair prices when you see them” to “hold memory purchases and watch spot.”
Burry has been early before, and being the most famous bear in the room is a business model as much as a thesis. But his Micron put is the first time someone with his platform has aimed directly at the fabric of this shortage.
