The AI Chip Stack

Read time: 4 min

Where the Bottleneck Is, Capital Follows

This week, we examine a core investment thesis: the AI hardware supply chain's bottleneck predictably shifts over time. Understanding its current and future location offers the most durable advantage for investors. A significant deal this week directly illustrates this thesis: today we delve into the physical hardware enabling models.

The key insight is straightforward: in any rapidly growing industry built on physical infrastructure, one input consistently faces supply shortages, granting it pricing power. Early identification of this input, before it becomes consensus, allows to capture returns. The challenge lies in the bottleneck's mobility; it moves towards the input requiring the longest build time.

Over the past 4 years, this migration in AI hardware has followed a consistent pattern. Today's edition maps its historical path, current position, and probable future direction.

Watch the 2-min video overview:

[Sources: Reuters, May 10 · Fortune, May 12 · Bloomberg, May 13 · CNBC, May 13]

Key numbers: Final IPO price $185/share · $5.55B raised · Valuation $56B+ · 20× oversubscribed · Largest US IPO of 2026

Cerebras Systems (CBRS) begins trading on Nasdaq today, having priced at $185 per share on Wednesday evening. This was above its already twice-raised range of $150–160. The final price gives Cerebras a market valuation exceeding $56B, making it one of the largest US IPOs ever. The company raised $5.55B from 30M shares, led by Morgan Stanley, Citigroup, Barclays, and UBS.

The product is unique. Cerebras builds chips using an entire silicon wafer as a single processor (58 times the size of a leading Nvidia GPU) with memory integrated directly onto the chip. This results in AI inference up to 15 times faster than GPU-based systems for specific workloads. This is not a minor improvement; it represents a different architectural approach to a problem becoming increasingly critical as the AI industry shifts from model training to large-scale deployment.

Financially, Cerebras reported $510M in 2025 revenue, a 76% growth, and a 47% net margin — a rare profitability for a pre-IPO hardware company. A multi-year deal with OpenAI for 750 megawatts of compute capacity and a binding term sheet with AWS provide near-term revenue visibility.

However, two risks warrant direct attention. First, customer concentration: nearly 90% of Cerebras' revenue comes from two customers. Second, and more specifically: despite 76% headline revenue growth, Cerebras' US domestic business shrank 34% year-on-year. All growth originated from international bookings concentrated in Abu Dhabi, stemming from the same G42 partnership that previously triggered a US national security review and caused the company to withdraw its initial IPO attempt in 2024. Although the national security review has since been cleared, the revenue geography presents a structural vulnerability obscured by the headline figures.

This geopolitical aspect directly relates to the investment thesis. Cerebras' architecture — memory on-chip, eliminating HBM requirements — effectively bets that the current hardware bottleneck is an investable problem. By circumventing the scarcity of specialized memory through direct integration, the constraint's impact diminishes. This is the deeper logic behind this IPO, beyond its headline valuation.

Investment note: At $185 per share, the valuation implies roughly 95× trailing revenue. This is a price-to-perfection multiple that leaves no room for execution error. The 20× oversubscription and three successive price increases signal genuine institutional demand. However, history is instructive: Figma, 2025’s most-hyped tech IPO, now trades approximately 40% below its listing price. The AWS deal is the most important de-risking signal, pointing toward a broader customer base beyond OpenAI and Abu Dhabi. Watch the lock-up expiry in November 2026 and the first public quarterly earnings as the two key inflection points.

The Cerebras IPO is today’s event. To understand its significance and the AI hardware investment landscape over the next 12–24 months, one must comprehend the supply chain it operates within.

AI hardware is constructed in layers, from raw materials and silicon wafers to chip fabrication, specialized memory, advanced packaging, servers, and ultimately, cloud platforms and applications. At any given moment, one layer acts as the binding constraint, where demand outstrips supply and pricing power concentrates.

Over the past 4 years, this migration has followed a clear path:

2022–23: The constraint was GPU chips themselves. Nvidia’s H100 faced queues stretching to 52 weeks. This was resolved when TSMC, the world’s largest chip manufacturer, dedicated more production lines to AI chips.

2024: The constraint shifted to advanced packaging — the specialized process connecting memory chips to GPU logic. Without this step, even finished chips cannot be assembled into functional AI servers. Lead times extended beyond 50 weeks.

2025: The constraint moved again, this time to specialized high-bandwidth memory (HBM), which sits alongside the GPU and feeds it data rapidly. Three global companies produce this memory, and supply ran short.

2026 — Now: Both of the last two constraints remain active. The newest generation of HBM, HBM4, is fully sold out before mass manufacturing. SK Hynix, the world’s largest supplier, publicly confirmed its entire 2026 production is committed, with Nvidia reserving approximately 70%. Samsung, the second-largest supplier, has not yet met technical qualification for shipping. Micron, the third, is still ramping up. Meanwhile, advanced packaging capacity for HBM4 assembly remains oversubscribed through at least mid-year, despite TSMC nearly quadrupling its output over two years.

Both memory and packaging constraints will ease as new factories are built. However, meaningful relief is 12 to 24 months away. Until then, pricing power resides in these two layers.

The more crucial question is what follows the easing of memory and packaging constraints. The answer is already apparent: power and electricity.

The four largest tech companies (Microsoft, Google, Meta, and Amazon) are collectively spending $485B on AI infrastructure in 2026 alone. This spending accelerates power demand at a rate the electrical grid was not designed to handle. In the US, the queue to connect new facilities to the electricity grid includes over 2,600 gigawatts of pending projects, with average waiting times exceeding 5 years.

As memory and packaging constraints ease through 2027, the bottleneck will shift to electricity. We covered early private-sector responses to this problem in last week’s edition — ocean-based platforms, space solar, and rocket-factory data centers. These investments are not exotic; they are rational responses to a predictable supply constraint not yet fully priced by the market.

The AI value chain is not a single trade but a rotating one. Companies with pricing power today — SK Hynix in memory, TSMC in packaging — will not necessarily hold it in 2027. The Cerebras IPO occupies an interesting position in this rotation: its architecture explicitly aims to reduce dependence on the current bottleneck by eliminating the need for external specialized memory. If successful, it is both a beneficiary of the current AI buildout and a structural challenger to a layer currently commanding the highest margins.

Portugal's leading enterprise AI gathering: 30+ keynotes, workshops, and panels across three stages covering AI strategy, engineering, and the future of work. Speakers include senior executives, founders, and policymakers. Early bird pricing (€405) expires tomorrow (May 15).

"The fox knows many things, but the hedgehog knows one big thing." — Archilochus, Greek poet, 7th century BC

That’s it for this week.

The CLNM Capital

Subscribe