@AyalinCreator on X
Agents need CPUs, so chip brains beat GPU bets
AI agents need as many server CPU cores as GPU power, so server CPU makers will win more than pure GPU plays.
Result
If held over the last 3 months (Aug 18, 2026 – Nov 17, 2026): −6.5%. Over the same days: Chip stocks −20.3%, S&P 500 −2.8%.
What it holds
- Server CPU kings, 38% of the money, −19.3% over 3 months: AMD, INTC (Intel). AMD and Intel sell the server processors that agents will need far more of.
- Arm and custom CPUs, 24% of the money, −25.7% over 3 months: ARM (Arm), GOOGL (Google). Cloud giants and Arm designs are packing racks with CPU cores too.
- The chip factory, 10% of the money, −11.0% over 3 months: TSM (TSMC). Whoever builds these new 2nm chips gets paid for every core.
- Pure accelerator bets, 24% of the money, +33.1% over 3 months: against NVDA (Nvidia), against CRWV (CoreWeave), against AVGO (Broadcom). If CPUs catch up in demand, companies built only on GPU-style chips lose their edge.
- Cash, 5%.
The plan behind it
The portfolio buys the companies that make server CPUs and the chips and clouds around them. AMD and Intel are the two big server CPU makers. Arm and Google make the other kind of server CPU, and TSMC builds the chips. It shorts the companies that depend most on GPU-style accelerators, because the take says CPUs will catch up in how many get bought. The longs and shorts are paired, so the portfolio can win even if AI spending only shifts toward CPUs.
The consensus narrative spent the last two years declaring the central processor dead in the water, reduced to a glorified traffic cop feeding data to clusters of GPUs. That thesis completely misses how autonomous agentic AI actually runs in production. Inference models generate tokens, but agents execute decisions. Every recursive tool call, code sandbox execution, memory retrieval, and multi-agent coordination loop lives directly on the CPU. The industry is realizing that scaling agentic intelligence does not just demand raw accelerator compute, it requires an unprecedented expansion of processor cores. Recent data highlights why server CPU shipments surged twenty percent year over year to 8.13 million units in the third quarter. Industry projections indicate agent ic workloads could quadruple data center CPU core density from 30 million to 120 million cores per gigawatt . That shifts the historical hardware balance, pulling CPU to GPU deployment ratios back toward parity and expanding the server CPU market toward 200 billion dollars by 2030. Hardware architecture is pivoting rapidly to meet this pressure. AMD just rolled out its sixth generation EPYC Venice processors on TSMC 2nm silicon, packing up to 256 cores and 512 threads into a single socket alongside a gigabyte of L3 cache. Meanwhile , Intel is preparing its 18A-P Nova Lake silicon with massive cache configurations to eliminate latency choke points between threads. Even custom Arm designs like Nvidia Vera and Google Axion are crowding server racks with high-density execution units. The future of AI infrastructure is not just a monolithic cluster of accelerators burning through gigawatts on training runs. It belongs to dense, cache-heavy processor cores capable of juggling millions of autonomous workflows without collapsing under orchestration overhead.
What would prove it wrong: Server CPU sales stall while GPU sales keep growing much faster, which would show agents don't need extra cores.
FAQ
How do I bet that AI agents need as many server CPU cores as GPU power, so server CPU makers will win more than pure GPU plays?
This tested idea holds AMD, INTC (Intel), ARM (Arm), GOOGL (Google) and TSM (TSMC) and bets against NVDA (Nvidia), CRWV (CoreWeave) and AVGO (Broadcom), with 5% in cash. Make a copy to run it on practice money first, at real prices.
What would prove it wrong?
Server CPU sales stall while GPU sales keep growing much faster, which would show agents don't need extra cores.
Can I make one from my own idea?
Yes. Write what you believe in plain words at https://f.xyz/create. fxyz asks a couple of questions, builds the portfolio and shows how it would have done over the last 3 months.
Can I turn a post on X into a theory?
Yes. Reply to the post and tag @fxyzai. It tests the view in the post and replies with the 3-month result and a link.
What can it trade?
Stocks (Tesla, Apple, Nvidia), crypto (Bitcoin, Ethereum, Solana), pre-IPO companies (OpenAI, Anthropic), indices (Nasdaq 100, S&P 500), commodities (gold, silver, oil, copper) and currencies (EUR/USD, USD/JPY). One theory can mix them, and bet for or against any of them.
How does it react to news?
On each check it reads the last two days of news on its biggest holdings and says whether the idea still holds. It can make one small change at a time: trim or drop one holding by up to 5 points. On autopilot the change waits 10 minutes so the owner can skip it; otherwise it asks first.
How do I run it with real money?
For real money, tap Make it mine, open your copy and choose Use real money, with an exchange connected: Hyperliquid, Lighter, Lighter (Robinhood), Backpack, Pacifica or Binance. The money stays in your own account there. fxyz can place trades, but it can never withdraw.
Can the AI get it wrong?
Yes. fxyz's AI turns the idea into holdings, and it can get that wrong: add a coin or stock that doesn't fit, or miss one that does. Check the holdings before you put money in. In your own copy, ask its chat to add or drop any of them.
What do the 3-month numbers mean?
What $100 would have done holding this mix over the last 3 months, at real prices. It is a what-if, not a record: past or simulated results are no promise of future ones, and nothing here is investment advice.