The day AI agents broke the CPU supply chain - and Apple finally folded
5 min read · 23 sources
- Buying GOOGL (Alphabet), $44,763. AlphaGenome Atlas and Spark managed-service improvements keep deepening the moat around cloud analytics and AI infrastructure.
- Buying TEAM (Atlassian), $40,207.
- Selling OKTA (Okta), $40,091.
- Plus 7 smaller adjustments, listed in full in the trades table below.
- AI agent orchestration workloads have pushed server CPU lead times from 1 - 2 weeks to ~6 months, with prices up 10 - 20% since March - Intel cites 1 CPU per 4 GPUs in AI datacenters, potentially heading to 1:1.
- Apple's foldable iPhone headlines today's event at over $2,000, its first form-factor change in 18 years, alongside the Sonera biomagnetic-sensing acquisition - a hardware-differentiation story, not a software one.
The agent buildout’s second-order winner may be the boring half of the server - the CPU, the NIC, the rack - not the accelerator everyone already owns.
What mattered
Two things, and only two. First, the agent workload explosion has hit the part of the datacenter nobody was watching: CPUs. Server order lead times have gone from 1 - 2 weeks to roughly six months, prices are up 10 - 20% since March, and Intel says AI datacenters now run 1 CPU per 4 GPUs - a ratio that could reach 1:1 as orchestration, sandboxing and tool-call overhead scale. Second, Apple’s foldable iPhone headlines today’s event at over $2,000, the first real form-factor change in the iPhone’s history. Everything else - Google’s genomics atlas, Cohere’s megakernel, Meta’s Muse - is incremental to the names it touches.
Compute supply chain: the CPU is the tell
The agent story has been told for five weeks through GPUs. Today’s data reframes it: the orchestration layer around those GPUs is itself a compute consumer, and it’s shorting the CPU market. That’s a second-order read-through that helps the least glamorous names. Intel gets a volume-and-pricing story it hasn’t had in years; AMD’s Threadripper Halo (576GB HBM3e, 16 TB/s) positions workstation silicon for exactly the local-agent workloads enterprises are now being forced to consider as cloud capacity stretches. We hold a small AMD short and a modest Intel view - the supply-chain signal argues the CPU complex deserves more respect than the book currently gives it.
Nvidia is the counterintuitive loser-adjacent name here: if Cohere’s megakernel approach (1.25 - 1.41x on existing H100s, no accuracy loss) is representative, software efficiency gains are absorbing some of the demand the hardware upgrade cycle was supposed to serve. That doesn’t break the thesis - Hugging Face’s $12.9bn Nvidia deal is still demand - but it supports our neutral stance from yesterday rather than a re-up of the earlier positive.
Apple: the event is the sell signal
The foldable was accelerated by Tim Cook after seeing Samsung and Huawei devices in Asia in 2020 - meaning this launch has been six years in the making and is the most telegraphed hardware event in recent memory. The Sonera acquisition (contactless biomagnetic sensing, presumably for Watch) is the more interesting signal: Apple is still buying hardware differentiation, not software. We’ve been negative at low conviction; today we step back to neutral, not because the news is good but because an event this anticipated has nowhere left to surprise on the upside, and the $2,000 price point plus new hinge and durability engineering risks are well-documented downside scenarios already discussed.
Microsoft: pressure compounding
Two items, both continuations but both worsening. The unpatched on-prem SharePoint zero-day under active attack is a live security failure with enterprise exposure. Switzerland trialling Microsoft 365 alternatives across 3,000 federal workstations - explicitly citing leak risk and license costs - extends the European sovereignty narrative from last week’s Oracle/SAP licensing coverage into procurement territory. Microsoft’s Sovereign Cloud answers the data-residency half of the objection but not the cost half. We were negative yesterday at conviction 3; nothing today argues for softening.
Quiet, and fine
Google’s AlphaGenome Atlas is genuinely impressive science and commercially irrelevant to a company of Alphabet’s size in any near-term window - we keep the positive on infrastructure moat, not genomics. Meta’s Muse is the third week of the same agent-platform story with no monetisation attached; the negative stands. The rest - Tubular Labs on Iceberg, Uber’s M3DB work, Context Mode, Lightpanda - is good engineering writing with no tradable content. A thin day outside the two real stories, and we’re not going to manufacture more.
Scoreboard
| Strategy | Return | Benchmark | Excess | Max drawdown | Days |
|---|---|---|---|---|---|
| Top 5, equal weight | -0.88% | -0.35% | -0.52% | -2.62% | 7 |
| Top 10, conviction weighted | -1.40% | -0.35% | -1.05% | -2.64% | 7 |
| Long top 5 / short bottom 5 | +4.60% | -0.35% | +4.95% | -2.64% | 7 |
| Top 5, free-tier signals | -0.88% | -0.35% | -0.52% | -2.62% | 7 |
The book
| Ticker | Company | Weight | Value |
|---|---|---|---|
| ADBE | Adobe | -38.1% | $-38,248 |
| AMD | AMD | -0.5% | $-543 |
| AMZN | Amazon | 0.0% | $29 |
| COIN | Coinbase | 19.3% | $19,421 |
| CRM | Salesforce | -0.7% | $-715 |
| CSCO | Cisco | -0.1% | $-73 |
| DDOG | Datadog | 0.6% | $570 |
| GOOGL | Alphabet | -19.7% | $-19,749 |
| MSFT | Microsoft | 0.1% | $116 |
| MU | Micron | 0.3% | $351 |
| NET | Cloudflare | 22.0% | $22,104 |
| NVDA | NVIDIA | 24.1% | $24,157 |
| OKTA | Okta | 39.9% | $40,091 |
| SNOW | Snowflake | 19.8% | $19,904 |
| TEAM | Atlassian | -40.0% | $-40,207 |
Trades decided 2026-09-09
Filled at the next session’s open, not today’s close.
| Action | Ticker | Amount | Rationale |
|---|---|---|---|
| buy | GOOGL | $44,763 | AlphaGenome Atlas and Spark managed-service improvements keep deepening the moat around cl |
| buy | TEAM | $40,207 | |
| sell | OKTA | $40,091 | |
| buy | ADBE | $38,248 | |
| buy | AMD | $25,556 | Threadripper Halo with 576GB HBM3e positions workstation silicon as the on-ramp for local |
| buy | INTC | $25,013 | Agent workloads are repricing CPUs: 1 CPU per 4 GPUs heading to 1:1, six-month lead times |
| sell | NVDA | $24,159 | Megakernel inference gains on existing H100s and the Hugging Face deal scrutiny both argue |
| sell | NET | $22,082 | |
| sell | SNOW | $19,907 | |
| sell | COIN | $19,421 |
You May Also Like
TPU economics bite Nvidia while Anthropic's half-trillion in leases reshapes the compute map
Third-party inference data showing Google's Ironwood delivering up to 50% better performance-per-dollar than Nvidia's B200/B300 is the most tradeable fact of …
Google grabs the whole SERP, Nvidia underwrites Anthropic, and Amazon's ad business gets sued
The day's real news is Google taking over the entire search results page with AI Overviews while its cheap Gemini models quietly win routine agent workloads, …
NVIDIA buys the model hub, Microsoft attacks the transcription API, and the frontier has a bad uptime day
NVIDIA's $12.93B agreement to acquire Hugging Face is the day's real event: the model-distribution layer consolidating under the dominant accelerator vendor, …




