This group makes the engines of artificial intelligence: GPUs (graphics processing units — chips originally built for video games that turned out to be ideal for the parallel math behind neural networks) and AI accelerators (custom chips, also called ASICs — application-specific integrated circuits — purpose-built for one job: training and running AI models). The unit of demand is raw compute, measured in FLOPs (floating-point operations per second — how many math calculations a chip can do). Every chatbot answer, every training run, every AI agent ultimately rents time on one of these chips. The designers are NVIDIA (NVDA), AMD (AMD), Broadcom (AVGO) and Marvell (MRVL) for custom accelerators, plus the in-house chips of Google (TPU — Tensor Processing Unit) and Amazon (Trainium). None of them runs its own factory — they design the chips and pay TSMC (Taiwan Semiconductor Manufacturing Company, the contract chip-maker) to manufacture them.
On the signals in the grounding material, demand currently exceeds supply: chips are reported sold out years in advance and lead times run 6-12 months. Supply is limited not by money but by two physical chokepoints — TSMC's advanced packaging (CoWoS) and high-bandwidth memory (HBM) — which take years and rare know-how to expand. In money terms, the market currently pays a large multiple of current sales for the leaders (commonly high-single-digits to ~20x revenue est.); arithmetically, a price set at that multiple already prices in years of continued fast growth. Whether the demand-over-supply gap stays open long enough to justify that price is for the reader to judge — this sheet states the facts and the arithmetic, not a verdict.
The product is a finished accelerator package: a logic die (the compute brains) bonded together with several stacks of HBM (high-bandwidth memory — fast memory placed right next to the chip so data doesn't starve the compute cores) on top of a silicon interposer (a base layer that wires the pieces together), then assembled using advanced packaging. One modern training chip is itself a small assembly of separately manufactured parts.
The unit sold is the chip — but the way it is actually consumed is as compute: FLOPs delivered over time. Buyers care about performance per chip, performance per watt of electricity, and performance per dollar, because a data center is constrained by power and budget as much as by chip count.
How the money is made, in plain cash terms:
Where demand is today (grounded): The scan describes accelerator silicon as "THE bottleneck" of the AI build-out — "every dollar of AI capex starts with accelerator silicon" (capex = capital expenditure, the money companies spend on long-lived equipment like chips and buildings). NVIDIA's Blackwell and next-generation Rubin parts are described as "sold out years ahead." Demand is currently constrained by supply, not the other way round — buyers would take more chips than can be made.
Source: 500-stocks scan, "GPU & AI Accelerators" sub-section (/Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/02-semiconductors.html).
Who the buyers are: a small number of very large spenders — the big cloud platforms (Microsoft, Google, Amazon, Meta, Oracle), a wave of "neoclouds" (GPU-rental specialists like CoreWeave that buy chips and rent out compute by the hour), AI labs (OpenAI, Anthropic and others, usually buying through a cloud), plus governments and large enterprises. This concentration matters: a handful of customers' capex budgets drive most of the demand, so the group's revenue is geared to a few balance sheets.
Forward demand (forecast — AGI lens): Reasoning from the premise that AGI is arriving, demand has two compounding legs. Training demand (the compute used to build a model) rises because each model generation has tended to need roughly an order of magnitude more compute, and recursive self-improvement (AI helping design and train the next AI) could multiply this further. Inference demand (the compute needed to actually run finished models for billions of users and autonomous agents) eventually exceeds training in this view, because it scales with every query and every agent action, not just with the occasional training run. The scan calls inference "the recurring revenue layer of the AGI stack." This paragraph is a forecast, not a contracted fact.
✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:
Source: AGI demand logic and "inference dwarfs training" framing from the 500-stocks scan (GPU and AI-Inference sub-sections). Market-size and growth figures: general knowledge, not live-verified est.
The binding constraint in the grounding material is not chip design or money — it is two physical inputs:
Source: 500-stocks scan, "Advanced Packaging (CoWoS)" and "Memory: HBM, DRAM & NAND" sub-sections. HBM growth/ASP magnitudes flagged est. (quoted from the scan, not live-verified).
Why supply can't just be bought: leading-edge fab capacity (fab = chip fabrication plant), CoWoS lines, and HBM production all take years to build and depend on scarce engineering know-how and a small set of equipment suppliers. Capital is available; the physical capacity and skills are the limiting factor, which is why lead times stay at 6-12 months even with heavy spending across the industry.
Market-share structure (who controls supply):
Source: company line-up from the 500-stocks scan; share percentages from general knowledge, not live-verified est.
On the observable signals in the grounding material, the product is short (demand exceeds supply): lead times of 6-12 months, next-generation parts reported sold out years ahead, CoWoS and HBM capacity expanding fast yet still reported unable to keep up, and pricing power described as strong (HBM alone reported at ~5x commodity memory ASP est.). These signals are the opposite of a market with spare capacity.
| Signal | What it shows | Read |
|---|---|---|
| Lead times | 6-12 months | Short (demand > supply) |
| Forward bookings | Blackwell / Rubin reported sold out years ahead | Short |
| CoWoS packaging | Tripling capacity, reported still unable to keep up | Short (tightest valve) |
| HBM memory | Demand +100%/yr est., ~5x ASP premium est. | Short |
| Pricing trend | Premium pricing reported holding | Short |
Source: 500-stocks scan GPU, CoWoS and HBM sub-sections; HBM growth/ASP figures flagged est.
When could it flip to oversupply (forecast): the gap closes if either side moves. Supply side — once TSMC's CoWoS and the three HBM makers' multi-year expansions all land at once, raw capacity could overshoot a demand wobble. Demand side — the risk is concentration: most demand comes from a few hyperscalers' (the largest cloud operators') capex budgets, so a pause in their spending (a "digestion" period after over-ordering) could create a glut quickly, as has happened in prior chip cycles. Reasoning from AGI arriving, the structural demand trend points up for years; on that view the near-term flip risk is a cyclical air-pocket from over-ordering rather than a permanent end of demand. Both timing and magnitude are unknown — this is a forecast, not a contracted fact.
| Company | What it makes here | Exposure to this product | Rough size est. | Position / edge |
|---|---|---|---|---|
| NVIDIA (NVDA) | Merchant GPUs + full racks + networking | Dominant; data-center compute is the large majority of revenue | Multi-trillion mkt cap est. | ~80-90% share est.; CUDA software lock-in |
| AMD (AMD) | Merchant GPUs (MI-series) + CPUs | Mixed; AI GPU is a fast-growing minority of revenue | Hundreds of $B est. | #2 merchant GPU; the main alternative |
| Broadcom (AVGO) | Custom AI ASICs + networking chips | Diversified; AI is a large and growing slice, not the whole | ~$1T+ mkt cap est. | Co-designs Google/others' custom chips; networking strength |
| Marvell (MRVL) | Custom AI ASICs + data-center connectivity | More concentrated on data center than AVGO; AI is a major driver | Tens of $B est. | #2 in custom ASIC; Amazon Trainium and others |
| Intel (INTC) | CPUs; AI accelerators (Gaudi) trailing | Small slice of AI accelerator demand | ~$100B-ish est. | Not a share leader in AI compute |
| Alphabet (GOOGL), Amazon (AMZN) | In-house chips (TPU, Trainium) | Tiny % of their revenue; built to cut their own NVIDIA bill | Multi-trillion est. | Buyers who became makers; demand pull-back risk for merchants |
Source: company list from the 500-stocks scan GPU sub-section; sizes, shares and revenue-mix all general knowledge, not live-verified est.
In plain money terms, here is what an owner pays today for a claim on the future demand-over-supply gap, stated as arithmetic rather than a judgment:
Source: fabless / cash-generation structure is well-known filing fact; sales-multiple ranges and crash-drawdown note are general knowledge / project notes, not live-verified est.
Factual pointers for where a company-level deep-dive would be most informative, by type of exposure (this is a pointer to where the information is richest, not a recommendation):
Source: as listed above. Forecasts (forward demand, oversupply-flip timing) are explicitly labelled forecasts in the Demand and Gap sections.