Hyperscalers are quietly turning semis into utilities

Hi {{first_name|Investor}} -

The semi industry has been priced as cyclical for forty years. A few lines buried in Intel's Q&A last Thursday suggest that frame is starting to age.

An analyst asked Lip Bu Tan about a new long-term agreement with Google. Lip Bu confirmed the deal — Xeon CPUs, some ASIC volume, multi-year. CFO David Zinsner added the part worth highlighting: "Most of these agreements are structured with volume and pricing, and they are usually somewhere between three and five years."

Then he said something even more important: "In some cases, customers want to keep that confidential, and we respect their desire to maintain confidentiality, so some of them we just did not announce."

There are more of these contracts than what Intel disclosed publicly. Across the AI infrastructure complex, hyperscalers are quietly locking in chip supply for half a decade at a time.

This is a much bigger story than one Intel-Google deal. It changes how to read every semi name in your portfolio.

A Long-Term Agreement (LTA) is a multi-year contract that locks in both volume and pricing between a buyer and a seller. Unlike a typical purchase order — which is essentially spot-market buying with some forward visibility — an LTA commits the buyer to specific quantities and the seller to specific prices, usually over three to five years.

Until recently, LTAs in semis were rare. The industry was too cyclical, demand was too unpredictable, and neither side wanted to be locked in. Buyers worried about overpaying if prices fell. Sellers worried about leaving money on the table if demand spiked. So most volume moved on shorter contracts, with constant renegotiation.

That has changed. Hyperscalers now have enough demand visibility — driven by AI infrastructure planning — to commit years out. Semi companies have enough supply constraint — driven by leading-edge capacity scarcity — to demand the same. Both sides finally want the same thing: certainty.

What you're seeing in real time is the contractual plumbing of the AI buildout being rewritten. And it has cascading implications for how to value the companies on both sides of those contracts.

Semis have always been a cyclical industry. That's the conventional wisdom and historically it's been right. Boom-and-bust cycles every few years, driven by inventory swings and demand shocks. Investors apply lower multiples to cyclical businesses because the earnings aren't predictable.

LTAs change the math.

Once a meaningful portion of a company's revenue is contracted out for three-to-five years at fixed volumes and pricing, that portion of the business stops behaving cyclically. It starts behaving like a utility — predictable, recurring, and resistant to short-term demand shocks. The market hasn't fully re-rated semis for this yet, but the structural shift is happening.

The clearest public-market data point sits in Intel's own commentary: Microsoft's "remaining performance obligations" — RPO, the accounting term for revenue customers have already legally committed to — are above $290 billion. That's not a forecast. That's contracted future revenue, sitting on the balance sheet, waiting to be recognized as the work gets delivered. Microsoft isn't a pure semi name, but the same dynamic is showing up across the supply chain.

For a portfolio aimed at compounding toward work-optional, this is structurally important. Contracted revenue compounds differently than cyclical revenue. It deserves a different multiple, a different position size, and a different mental model.

If LTAs are quietly transforming semi cash flow profiles, the question is how do you spot it? Here's what to look for on the next round of earnings calls.

One — Remaining Performance Obligations (RPO). This is the single most important number you can track. It sits in the financial statements and shows revenue under contract that hasn't yet been recognized. Watch the direction: rising RPO with rising contract duration means more of the business is becoming utility-like. Falling RPO is the opposite signal.

Two — Named LTAs versus confidential ones. When a company announces an LTA with a named hyperscaler customer (like Intel-Google), that's hard signal — both parties wanted the disclosure for strategic reasons. But the unnamed contracts matter just as much. Pay attention when management uses phrases like "multiple long-term agreements" or "significant contracted volume" without naming the counterparty. The book is bigger than what's disclosed.

Three — Contract length and pricing structure. Three years is good. Five years is better. Contracts that lock both volume and pricing — rather than volume alone — provide the strongest cash flow predictability. Listen for whether management discloses the pricing mechanism. Vagueness is usually the answer to a question worth asking.

Four — Repeat customers. Hyperscalers signing second and third LTAs with the same supplier are demonstrating real conviction. One contract is a transaction. A pattern is a moat.

Apply this lens across the AI infrastructure complex and several names look different than the conventional narrative implies.

Foundries. TSMC (TSM) has been the original LTA story for years — hyperscalers and chip designers reserving capacity at leading-edge nodes far in advance. Their CoWoS packaging contracts in particular have shifted from year-to-year to multi-year commitments. The cyclical multiple investors still apply to TSM understates how much of the business is now contracted.

ASIC designers. Broadcom (AVGO) designs Google's TPUs and increasing custom silicon for other hyperscalers. These are inherently multi-year programs — a custom chip takes 18–24 months to design and tape out, and the customer commits to the full program before the first wafer is fabricated. That's an LTA in everything but name. Marvell (MRVL) has similar dynamics with Amazon and Microsoft programs.

x86 incumbents. Intel (INTC) is the most direct example given this week's disclosures. AMD (AMD) is likely seeing similar contract structure but discloses less specifically. Worth watching their next earnings call for RPO trajectory.

Memory. Micron (MU) and the Korean memory makers (Samsung and SK Hynix) have moved a meaningful portion of HBM (high-bandwidth memory) volume to multi-year supply agreements with hyperscalers. This is the most under-appreciated transformation of the entire group — memory has historically been the most cyclical segment of semis, and HBM LTAs are quietly making that less true.

The pattern isn't universal — pure-play GPU names are still mostly transactional, and lower-end semis remain cyclical. But across the leading-edge AI complex, LTAs are now the dominant contract structure for major customers. That's a structural shift worth pricing in.

Note: I hold positions

Three caveats worth applying before getting too excited about the contracted-revenue reframe.

One — Contract enforcement is more fragile than it looks. If AI monetization disappoints, hyperscalers will push to renegotiate. Most LTAs have penalty clauses, but enforcement is costly, slow, and damages future commercial relationships. A renegotiated contract is a worse contract — same supplier, smaller volumes, weaker pricing. The cleanest contracts on the books today only matter if both parties still want them to.

Two — Concentration risk inside the LTA. If 40% of a semi company's revenue is contracted to two hyperscalers, you've reduced cyclicality but added counterparty concentration. That's a different risk, not a smaller one. Worth scrutinizing the customer concentration footnotes.

Three — Pricing stickiness in a falling-cost environment. LTAs that lock in pricing at today's elevated levels look great today but could become competitive disadvantages if input costs fall sharply. Less likely in the current environment, but worth flagging.

This is the part I keep coming back to. A portfolio built for compounding toward financial independence isn't trying to win every quarter. It's trying to own businesses whose cash flows are predictable enough to compound through multiple cycles.

LTAs are doing exactly that to a meaningful slice of the semi industry. The companies with the largest contracted books look less like cyclicals and more like infrastructure compounders. The market hasn't fully repriced them yet. That gap is where the lens earns its keep.

As always: I won't tell you what to buy. I'll sharpen the lens you use to look.

Stay disciplined, Koh

Disclaimer: Nothing in this newsletter constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed above. Always do your own research and consult a licensed financial advisor before making investment decisions.