Everyone Hates AI Right Now: Four Stocks That Stay Bulletproof

Four big stories landed in the same week warning that AI is too powerful and must slow down — a model trying to flee testing, another blackmailing developers not to be shut off. If you hold Nvidia or a 401(k) full of it, the question is simple: does a brake mean spending stops? The video says not a single dollar has — Anthropic signed $45 billion in compute commitments 17 days before its slowdown essay, and two days after, Broadcom ’s CEO told TV his demand forecast for AI compute infrastructure had not budged a bit.

Why the Tap Stays Open Even as the Debate Slams the Brakes

The gap between narrative and cash is physical. A new Nvidia Blackwell rack draws 140 kilowatts , more than twenty old A100 servers combined, and needs liquid cooling with pipes under the floor. Fewer than 4% of U.S. data centers can take one. Microsoft ’s Satya Nadella said it plainly in November 2025: the problem is not chip supply, it is not having warm shells to plug into. A warm shell is a finished building with power and cooling already running — so the new rack goes into a new building and the old building keeps running its old chips, because the alternative is an empty room still paying its lease.

History helps size the boom. Every big buildout is measured as spend in a year divided by GDP that year. U.S. housing at the 2005 peak took 6.6% of the economy, railroads in the 1870s ran ~5% . AI sits 1–2% in 2026 and, if Goldman Sachs is right, heads toward 3% by 2028 — a $1 trillion-plus AI investment wave with $1.7 trillion already signed by Amazon, Google and Microsoft combined, about three years of buildout at the 2026 pace. Google alone pushes 330 times more AI through its data centers than two years ago, and Jensen Huang reckons an AI agent doing a job uses 15–100 times the compute of a person.

The Most Real U.S. Buildout in a Generation — Measured Against GDP

The bill is already written; the question is who collects. The video splits the data center into four buckets: the processor that thinks , memory that holds the conversation , light that moves answers , and the owner who rents the finished machine by the hour . Working memory grows with every word — a chat as long as a 300-page book needs ~40 GB just to remember itself. A Hopper chip shipped in 2022 holds 80 GB , and the model itself must live there too. When the model is bigger than the chip, the conversation spills into a rack-scale pool of separate memory nearby — a purchase data centers barely made before. Micron told investors in June that customers are doing exactly that.

When that memory sits in another rack, answers must commute, and copper cannot carry fast enough — so they ride light through glass fiber . That is the optics bucket; companies like Lumentum make the lasers turning electrical signals into light, and a new rack needs far more of them than an old one. The final bucket owns the finished computers and rents them out — an agent working a job for five straight days is renting the whole time. The thesis: do not guess which AI lab wins; own the parts none can win without , inside buildings that already have power and a lease.

Why Light Beats Copper

Are old chips scrap? Look at the rack. On stage in March 2025 Jensen Huang said once Blackwell ships in volume you could not give Hoppers away ; eight months later his CFO told investors every Nvidia chip ever sold and still plugged in is fully utilized , back to A100s shipped six years ago . Both are true because of the building. CoreWeave rents chips by the hour and in August 2026 signed an A100 contract to 2029 — pricing on those 2020 chips is way above a few years ago. In top markets, waiting for a grid connection is ~4 years , and the equipment to fix that is years out too — a multi-layer bottleneck no one talks about enough.

All of it depends on copper, and here the sponsor segment lands on Copper One Resources . In August copper topped $6.70 per pound on COMEX , up over 40% in 12 months — an all-time high. Billionaire Robert Friedland says there is no rational price for what you must have; Stanley Druckenmiller calls copper the tightest position he has ever studied . At record prices Copper One is down ~75% in 2026 after junior miners sold off hard when the Iran war began in late February. It was one of few juniors raising money straight through, so it now holds its largest cash and working capital ever : market cap about CAD 14.8 million , CAD 10.4 million of it cash, leaving the market pricing three copper projects at about CAD 4 million — past-producer Majuba Hill in Nevada plus Red Roanda and Red Hill in British Columbia, with Roanda drilling now . Early-stage, real risk — do your own diligence.

The Heart of the Money: A 20-Million Calculation Signal

Story and entry are different things; every name here runs through a 20-million-calculation nightly system. Output is five rungs: buy a lot, buy, buy a little, hold, sell . It starts with fair value versus price — if value sits above price, it leans buy — then overlays price trend (climbing into a turn is preferred) and, crucially, willingness to sell — the model’s sell guidance has been accurate in seven of the last eight years . It also hunts for problems; when it finds one, it notches the signal down . That happens to three of the seven names here.

A new data center needs new chips; the hardest part to get is the memory bolted right next to the processor . Micron stands among merely three global producers able to build that HBM class — and there is not enough. Two years ago it kept 11 cents of each revenue dollar as operating profit; last quarter it kept 80 cents . Enrichment: Micron ’s Sept 23, 2025 Q4 FY25 release showed record revenue and a guide to >50% gross margin with ~$1.2B sequential growth , and The Register on Sept 24 noted it is close to selling all HBM it will make next year — both explain why the stock, though above its 200-day and just under its 50-day , screens as buy a lot on price but expensive only 2% of its 10-year history , so the system cuts to buy a little pending the Sept 30 report.

The same shortage hit file storage. SanDisk makes NAND flash holding what models read and became independent on Feb 21, 2025 when Western Digital spun it out . A year ago barely breakeven, last quarter it kept 78 cents per dollar as operating profit, and price sits under 8 times next year’s expected earnings — a PEG of 0.3 (price divided by growth, ≤1 is cheap). Price and trend also say buy a lot , but with 19 months as a public company there is no 10-year history to check against; that absence alone notches the signal to buy — no lean-in without a history anchor.

The same buildout pays optics, but the market prices it toward a sale. Lumentum makes the lasers turning electrical into light so chips can talk; a new rack needs far more than an old one. Two years ago losing money per dollar, now keeping 27 cents as operating profit — validated by FY26 Q1: $1.01B revenue, 27.8% GAAP operating margin versus FY25 Q4: $480.7M, 33.3% GAAP gross margin . Headlines said it missed in August , yet the stock rose 13% next day because the business kept doing its thing. Still, price may have run beyond the business: 195 times last year’s earnings versus a ~70 times norm. Even with a PEG under 1 , the signal is possibly sell some — business fine, price ahead of itself.

Next is Marvell , which designs custom AI chips big clouds build instead of buying Nvidia’s, and one customer already owns a piece of it. In August 2026 Google took the right to buy about 6.7% of Marvell , earning it by buying roughly $120 billion of chips through 2033 — CNBC on Aug 19 pegged the option pool up to $12.2B in shares , locking seven years of buying. The problem is price has locked seven years of growth: 80 times last year’s profit versus a 33 times norm — one of the strongest sell signals in the video.

The inference renters are the most beaten-up lately, interestingly while parts makers get paid upfront. Cerebras builds the biggest chip in the world — a single wafer-scale engine the size of a dinner plate , 4 trillion transistors, 900,000 cores per third-party specs; some buy the machine outright, most sign multi-year contracts — 70 cents of each dollar comes from those. It lost $450 million last quarter, mostly stock compensation at IPO that never left the building. Cash $8.6 billion , backlog $25 billion of signed orders versus under $1B sales this year (TechCrunch: $5.5B raised at $185/share IPO in May 2026 ). Price alone says buy a lot and it slides, but four months public means no history and fair value can only be built off Wall Street estimates — diversity of valuation collapses to one method, so the signal is notched down one to buy ; watch whether backlog converts on schedule or gets pushed.

The other renter bets on everyone else, including the old ones — CoreWeave , which signed that A100 deal to 2029 , renting Nvidia chips by the hour to labs and clouds that cannot build fast enough. Two years ago $400M a quarter , now $2.575B in Q2 2026, up 112% YoY with $104B backlog and customers signed for 4.2 gigawatts of power against 1.5 running — three times what it can deliver today . Risk shows in the balance sheet: burned $13.7B cash in the last 12 months against $7.6B sales , borrowing to build ahead of demand. The voiceover here is re-recorded — screen numbers are accurate where the lip sync slipped. The anchor ends with the name you already own. Nvidia : two years ago $30B in a quarter , now $96.2B in Q2 FY27, up 106% YoY, Data Center $89.0B up 117% , keeping 66 cents per dollar as operating profit versus 62 cents two years ago, yet supplying only ~70% of what customers ask by its own estimate. The cringe part: guaranteed up to $18B of customers’ data-center leases — if they cannot pay, Nvidia co-signs. Price today is 28 times last year’s profit versus a 52 times norm over five years , PEG 0.35 — cheapest growth on this list and the only name priced as if nothing strange is happening . Price says buy a lot , holds above its 200-day even after a soft week, history check clean — signal stays buy a lot at a $5 trillion market cap, with dollar-cost averaging the practical path if the step is held; if price sits on the sales step, everyone else is overpaying. The video’s final warning is the tripwire: if that opening $1.7T signed backlog starts to shrink, all buys get tossed; until then, labs can debate slowing down while the buildings and what is inside keep growing.