Stop treating 'NVIDIA alternative' as a slide. It is a 42x tax until the software is actually fast. If you are signing a multi-year GPU deal, the TPU externalization numbers are the negotiation chip. If you already bet the company on CUDA lock-in, you need a written answer for when a customer or a director asks why Google is cheaper per token.

NVDA: the CUDA moat print is supportive near-term and does not kill the multiple by itself. GOOGL: TPU as a for-sale inference product is the under-modeled call option — and the threat. AMD: a working image with 14.8–42x worse economics is not a win; it is a reminder that 'we run the model' is not the same as 'we make money running the model.' AVGO / custom-ASIC names catch a bid if more hyperscalers decide CUDA is optional for inference.

CUDA kernel and serving-stack talent stays scarce. 'We can port to ROCm next quarter' teams should not sleep well — the public scoreboard just said they are not close. TPU/compiler and inference-economics roles get pulled forward. Generic 'AI platform' headcount that cannot move a dollar-per-token number is the first line on a freeze list.

Your fully-loaded software engineer is not competing with last year's chatbot. They are competing with a $8 task that already ties the best coding agent. If your 2026 plan still staffs 'tickets in, PRs out' as a human factory, you are budgeting for a cost structure the vendors are actively collapsing. The surviving org chart is judgment, review, and production ownership — not prompt-and-paste.

ANTH / OpenAI private marks, MSFT (GitHub/Copilot), and any 'AI coding copilot' multiple now have a public price-per-task. Cognition just showed that mixing a frontier model with a cheap sidekick holds quality and cuts the bill. That is margin pressure on single-model agent SKUs and a demand tailwind for whoever sells the cheap sidekick tokens. It is also a slow bleed on IT-services billable hours.

Junior ticket-takers and PMs whose output is a diff are on the clock. Staff engineers who can take an agent from 90 to production stay expensive. If your manager cannot explain why a human should own a task that Fusion already scores 62 on, assume the task will not have a human owner by year-end. This is not a 2030 story.

A six-month model lock is now a competitive risk, not a convenience. Dual-source the stack (closed frontier + cheap open-weight) or accept that a competitor will do the same work at a fraction of your token bill. Open-weight at 18B active in the top 5 is the tell: the 'only the labs can do this' talking point is leaking.

Closed-model vendors are in a feature race that compresses pricing power. Cloud inference (CRWV and the hyperscalers) wins volume if units get cheaper and usage explodes — Jevons, if it shows up in actual tokens, not in a slide. Model-picker tools and gateways become the switching layer. Single-model ISVs get marked down the first time a customer asks 'why aren't we on last week's frontier?'

Prompt-only specialists decay with every Pareto move. Evaluation, eval-harness, and 'which model for which task' people get more valuable. If your job is being the in-house expert on one branded model, you have the same problem as a reseller of last quarter's SKU.

Your GPU PO is not a strategy. The queue that actually determines whether you get power, cooling, slides, substrates, and people to turn a wrench is already sold out. If you do not have a relationship with the third-derivative vendors, you are bidding leftover capacity. Modular is not a buzzword. It is how the industry is bypassing a labor shortage you cannot hire your way out of in Phoenix.

The crowded trade is still accelerators. The less-crowded cash-flow is rails, film, chemistry, wet process, modular electrical, and factory skids: 2049.TW (King Slide), 2802.T (Ajinomoto), 4971.T (MEC), 323280.KQ (Taesung), FIX (Comfort Systems), PWR (Quanta Services), VRT (Vertiv). These names can re-rate on a sold-out print and de-rate the second the buildout pauses. Citrini's warning applies here too: the technology can keep improving while these stocks go down if the market is already paying for 61 GW.

Electricians, modular assemblers, and factory techs are the real shortage. 'AI strategist' is not. Datacenter construction labor is being designed out (up to ~70% fewer on-site hours). If your employment is field installation of power and cooling, the factory skid is the competitor. If your employment is writing AI roadmaps with no procurement authority, the market already priced you as optional.

Treat HBM like EUV in 2018: a chokepoint with one-and-a-half real suppliers. Dual-source in slides is not dual-source in wafers. If a China-for-China memory stack is in your risk register, 25% yield means it is not a 2026 escape hatch. Budget KV-cache memory as a first-class opex line, not a rounding error on the GPU quote.

000660.KS / Hynix and MU catch the 'they can actually stack this' premium. Samsung foundry/HBM is a swing factor. CXMT is not a listed relief valve. MRVL is talking its book on memory-attached XPUs and may still be right. A political green light for CXMT/YMTC into Apple would be a gap-down event for incumbent memory and a political firestorm — it is a rumor, not a shipment.

Packaging, TSV, and HBM yield engineers are among the scarcest people in the industry. Generic DRAM process talent is not the bottleneck. If you are in a China-memory program promising HBM parity on a slide, the 25% yield print is the career risk, not the opportunity.

If Intel is a supplier, expect more price and less Small Core. If Intel is a foundry partner, 'evaluate' from AMD is not a PO. If Intel is your employer, do not wait for the all-hands. The DIGITIMES sourcing is rumor-grade — treat it as a scenario, not a filing — but the shape (price up, people down, server crowding PC) is the same movie they have been in. Have a 18A/14A tell you actually believe, or do not staff a turnaround you cannot see.

INTC: a price hike into a shrinking PC TAM is a margin patch, not a growth story. QCOM / MediaTek: Small Core EOL is an opening in IPC/IoT, not a PC franchise steal overnight. TSM: Intel sending more server to TSMC would be a win for TSMC and a confession from Intel. This is rumor until it hits a 10-Q or a customer guide.

Intel employees in PC Small Core, DCAI leftovers, and 'corporate' layers are the implied target. Hiring 'alongside the cuts' is not safety; it is a mix-shift. ARM-camp IPC/IoT design jobs are the offset, and they will not be in the same buildings.

If you are a mid-size ASIC customer, you are the volume Samsung can refuse. Budget a lost-node delay as a base case, not a tail. If you need US-sited wafers, Taylor being up is news; yield and who gets the capacity are the actual questions. Custom silicon with Qualcomm for a hyperscaler is now an announced product path, which means your 'we will just buy GPUs on the open market' plan is competing with someone else's captive stack.

005930.KS: pricing power at Foundry is the bull case; losing Qualcomm volume is the hole in it. QCOM: ASIC services to ByteDance and Amazon is a second engine besides handsets — and a geopolitical headache. TSM remains the default. TSLA/AVGO as Samsung foundry customers are the reason Qualcomm got pushed.

Foundry sales and capacity-planning jobs at the leading edge are leverage jobs. Backend 'we will find you wafers' broker roles shrink as the queue hardens. US-fab technicians at Taylor are a real hire print. If your company promised a 2026 custom chip that is not already in a foundry slot, the employment risk is the project, not the process node.

Hardware launches at this price are yield stories wearing a keynote. If you are an Apple supplier, golden-sample mode means your second-source status is a trap (building to someone else's frozen spec). If you compete with Apple, do not assume 5–6 million units hit the channel. If you run a consumer P&L, $2,000 is a different buyer — conversion will be worse than the last iPhone cycle even if the device is fine.

AAPL: a delayed or ugly foldable is not an iPhone franchise killer, but it is a narrative risk into October 16. 3376.TW (Shin Zu Shing), Amphenol, Foxconn, Lingyi: binary to hinge/backplate yield. Samsung Display still has the OLED. Fade the 'foldable supercycle' multiple until build reports stop saying golden samples.

Precision hinge and 3D-printed module process engineers are scarce. Cosmetic-industrial designers just got a public dragging. Foxconn line workers sit behind a parts gate they do not control. If your bonus is Duo unit volume, the hinge vendor's yield is your employment risk.

Separate the technology curve from the ticker. If your board pack still equates 'model eval up' with 'our AI investment is working,' you are managing a mood, not a business. Decide whether you are building on distributed models or betting on a regulatory moat. Those are opposite strategies and they will not both be right.

This is the stock-price section in one paragraph. NVDA, AVGO, the software-AI compounders, and the third-derivative industrials can all print worse while Fable/Astra/Fusion get better — if the market already paid for the buildout, if ROIC on capex is in doubt, or if regulation becomes the product. Citrini selling is not a short recommendation. It is a reminder that the people closest to the narrative are rotating the vehicle. Watch the new fund's actual book, not the goodbye post.

Research-shop talent follows the founder. If you were at the old firm, you already know. For everyone else: 'AI analyst' whose product is ticker enthusiasm is the job Citrini just mocked. Analysts who can hold 'models up, stocks down' in their head keep a seat.

If you are a CFO, the MSFT line is the new template: debt is fine when the incremental return is real, and only then. If you cannot show ROIC on GPUs, do not copy the debt. If you are renting compute, you are in the OpenAI/Anthropic shape — opex that can be repriced by the landlord. If you own the fleet, you have the Meta/Google problem: allocation fights inside the parent. Either way, 'we spent it' is not 'it earned it.' The Navier-Stokes/Codex story is a reminder that training-data provenance is now a legal and reputational line item, not an ethics sidebar.

MSFT: debt-financed AI capex is fine until ROIC misses; then it is a multiple compression story. ORCL, AMZN, GOOGL, CRWV: the landlords of the 20x compute spike. A pause in rental demand is their gap-down. OpenAI/Anthropic private marks assume the rental flywheel never jams. Watch 10-Q capex and debt footnotes, not keynotes.

Finance partnered with infra (people who can talk ROIC and rack density in the same sentence) gets hired. Pure 'growth at all costs' AI ops gets a ceiling the first quarter debt service shows up. Legal/compliance on training data just got another exhibit. If your lab's compute is rented, a landlord's price change is a layoff mechanism.