Every AI datacenter interconnect ends in a laser made on indium phosphide, and in 2026 the fabrication capacity for that laser — not GPUs, not HBM, not power — became the tightest link in the chain. The clearest evidence is what a customer did about it. On March 2, 2026, NVIDIA announced "strategic investments of $2 billion each" in Coherent Corp. and Lumentum Holdings, inside non-exclusive multiyear agreements that also include "multi-billion-dollar purchase commitments" and "access to future advanced laser and optical networking products and capacity." NVIDIA's stated rationale: optical interconnects and advanced packaging integration are "foundational to next-generation AI infrastructure." This piece reads the two suppliers' filings and earnings to answer four questions: what exactly was bought, why this material, whether the demand is measured or asserted, and why capital cannot fix it on the schedule the AI build-out wants. Every figure comes from company releases, SEC filings, or named secondary reporting; the arithmetic is ours and is labelled as such.

What did NVIDIA buy on March 2, 2026?

Three things, in one structure. First, purchase commitments — multi-billion-dollar, multi-year, non-exclusive: a demand floor under the expansion. Second, capacity rights — "access to future advanced laser component capacity," language that treats fab output as something to be reserved rather than bought spot. Third, equity — $2 billion per company. The equity's shape matters more than its size: Lumentum's fiscal-2026 annual report records an investment of approximately 2.9 million shares of Series A convertible preferred stock at $695.31 per share, providing net cash proceeds of $1,999.7 million. A preferred convert is patient capital — it does not demand a seat at the pricing table, it funds cleanrooms. Lumentum's CEO Michael Hurlston said the company would use it to "invest in building a new manufacturing facility to increase capacity"; Coherent's CEO Jim Anderson framed the deal as cementing Coherent's position as "a key enabler of AI datacenter infrastructure."

Scale it against what the suppliers were already spending. Lumentum's capital expenditure for fiscal 2026 (ended June 27, 2026) was $451.3 million, nearly double the prior year's $231.0 million. The NVIDIA investment of $2 billion is 4.4× the recipient's entire annual capex (ours). Coherent's fiscal-2026 capex intensity ran roughly 15.6% of its $7.12 billion revenue — about $1.1 billion (ours, derived from the reported ratio) — so NVIDIA's two checks together exceed the two companies' combined annual capital budgets. When a customer's investment exceeds the supplier's whole expansion budget, the customer is no longer buying parts. It is underwriting the factory.

The pattern has precedent in this cycle — NVIDIA has invested in datacenter operators and model developers — but this is one level deeper in the stack. Those investments bought demand for compute. This one buys inputs to compute: the laser chips that carry gradient and weight synchronisation between accelerators. Vertically integrating demand is a 2025 story; vertically integrating materials is a 2026 one.

Why is indium phosphide the bottleneck?

Because the move to 1.6 terabit transceivers multiplies laser complexity on a wafer process that cannot be rushed. Each 1.6T module in Lumentum's design uses four 400G-class EML (electro-absorption modulated) lasers, each doubling the per-lane data rate of the 800G generation. Doubling per-lane rate tightens the wafer spec — thinner tolerances on the epitaxial layers that make a laser emit at the right wavelength and modulate cleanly — which means lower yield per wafer even before volumes rise. A shortage of bare InP wafers or wafer fab capacity therefore restricts module deliveries directly, regardless of how many heat sinks, packages or fibres exist. Unlike silicon, InP is not a substitute game: it is the material that efficiently generates and amplifies light at fibre wavelengths, and both Coherent's and Lumentum's leadership identify it as the core of their datacenter franchise.

The demand signal is quantified, and it is extreme. Lumentum's quarterly revenue ladder reads: $480.7M (quarter ended June 2025) → $1.01B (June 2026) → $1.26B (September 2026), a 12-month multiple of roughly 2.6×, with the December quarter guided to $1.575–1.625 billion — sequential growth of about $350 million, which Hurlston noted is "greater than the quarterly revenue of the whole company just two years ago." Fiscal-2026 revenue was $3.01 billion against $1.6 billion the year before, +83%. Coherent shows the same curve: fiscal-2026 fourth-quarter revenue of $2.05 billion, +34% year over year (+42% pro forma), full-year revenue of $7.12 billion (+22.5%), and — the number that isolates the AI effect — datacenter and communications revenue up 59% over twelve months, from $1,018 million to $1,615 million, now 79% of company revenue. In November 2026, Lumentum announced a "second major NPO/CPO win, backed by purchase commitments exceeding $5 billion." Five billion dollars of committed purchases for one product line, at one supplier, in one quarter.

Against that, the supply side speaks in the present tense of shortage. Lumentum's annual report states that its operations and those of its clients "are experiencing" supply limits from tight capacity and parts shortages, which have "adversely affected" availability, delivery schedules and expenses. Hurlston, quoted by TrendForce on August 6, 2026, said the market imbalance in indium phosphide "now surpasses that of DRAM and NAND memory" — a comparison to the most famous shortage of the last memory cycle, made by the operator of the largest capacity. When the supplier of a $5 component says its scarcity exceeds DRAM's, the constraint is not a sales narrative; it is a yield curve.

Why doesn't capital fix it immediately?

Because compound-semiconductor fabs ramp on years, not quarters, and the money is arriving behind the demand curve. Lumentum nearly doubled capex to $451.3 million in fiscal 2026; net property, plant and equipment grew roughly 60%, from $726.4 million to $1,159.1 million; construction in progress stood at $377.4 million — and the annual report says management plans to commission it "over the next 12 months." The Greensboro, North Carolina site acquired in March for $38.0 million in cash adds land, buildings and staff, but management's own disclosures describe cleanroom bring-up and wafer-yield tuning as the slow phases. The arithmetic of the mismatch is simple (ours): revenue ran 2.6× higher over twelve months while the physical asset base grew ~60% — the excess is being monetised as pricing power and lead times, not shipped as product. Coherent is expanding capacity too, and its CFO tied capital allocation explicitly to "expand manufacturing capacity" for "the ongoing acceleration in customer demand."

One expansion has dates attached. Coherent's Sherman, Texas six-inch InP wafer fab — where NVIDIA's investment supports US manufacturing — broke ground on June 16, 2026, the same day Coherent announced a $50 million CHIPS Letter of Intent from the US Department of Commerce; industry reporting describes the completed line as multiplying site capacity about four-fold. That is the first structurally new capacity with a public timeline, and it lands after the 1.6T transition is already in volume ramp. This is why both companies' leadership talk about tightness persisting through 2026 and 2027 rather than easing with the next capex cycle: the constraint is not willingness to spend, it is the physics and cleanroom time of III-V wafer fabrication.

There is also a cautionary tale inside the numbers about reading GAAP headlines. Lumentum's fiscal-2026 results show a GAAP net loss of $6.9 billion — driven by a one-time, non-cash loss on debt extinguishment of $7.8 billion as convertible notes were equitised — while non-GAAP net income for the same year was $782.3 million and Q4 non-GAAP net income was $326.3 million. The laser business is printing money; the balance-sheet restructuring that accompanied the NVIDIA investment made the annual report look like a catastrophe. Anyone screening "unprofitable supplier" out of the supply chain read would have screened out exactly the wrong company.

Where does the geopolitical risk sit?

Upstream of both suppliers, in one jurisdiction. China refines roughly 70% of the world's indium — the metal InP wafers are grown from — according to US Geological Survey data cited in industry reporting. On February 4, 2025, Beijing imposed export licensing requirements on indium and indium phosphide; reporting since then describes permit processes adding delay to international shipments. The CEO of IQE, the Cardiff-based substrate maker, warned in September 2026 that access to InP substrates "is increasingly constrained by geopolitical factors" as her company reported AI-driven revenue growth. Industry reporting through 2026 describes substrate prices rising sharply since the controls — with spot-market premiums and buyers prepaying 30–50% to lock capacity years out — though the specific per-wafer price points circulating in press coverage are single-source figures we have deliberately not repeated here.

The US policy response is running on two tracks, both visible in the Sherman, Texas project: CHIPS funding for domestic InP capacity, and — according to Reuters reporting relayed by industry press — potential FCC restrictions on new Chinese-made optical transceivers entering the US market. Note what this combination does to the supply picture: if the demand shift away from Chinese module vendors accelerates, non-Chinese laser capacity gets tighter still, because the same fabs now serve a larger share of Western demand. Export controls plus reshoring incentives plus a demand surge is a squeeze on all three fronts.

What does the $4 billion actually price?

It prices wafer time, and it prices it through an option structure rather than a spot purchase. The preferred stock funds the cleanrooms; the purchase commitments guarantee the cleanrooms a buyer; the capacity rights give NVIDIA first call on output. Put the three legs against Lumentum's November disclosure — a second NPO/CPO win backed by purchase commitments exceeding $5 billion — and the structure reads plainly (ours): roughly $2 billion of patient equity is anchoring more than $5 billion of committed demand, i.e., NVIDIA paid about 40 cents of balance-sheet capital per dollar of secured supply. That is cheaper than paying spot prices for scarce modules later, and more reliable than forecasts, because the capacity is contractually reserved.

It is also the same pattern this forum has documented up and down the 2026 stack, one layer deeper each time. Capacity markets sell options on megawatts standing by ($333.44/MW-day in PJM's 2027/28 auction). Gas turbine sellers sell options on production dates (slots reserved to 2032, callable by the seller). GPU clouds sell commitments against list-price anchors. Now a hyperscaler chipmaker is buying options on laser wafer output — and, notably, it is the buyer doing the option-writing this time, pre-paying the premium in equity so the strike price later is availability rather than money. The scarce products in this build-out are never the parts on the invoice; they are the queue positions behind them.

For anyone planning AI infrastructure spend in 2027, the read is concrete. Optics prices and lead times will stay seller-set at least until Sherman-type capacity reaches volume and Lumentum's construction-in-progress converts to shippable lines. Module-level prices are the visible symptom; wafer-level contracts are the market. And the NVIDIA investment sets a template other platform companies are likely to copy: the next reservation of critical input capacity will probably also be written in preferred stock, because that is now the demonstrated price of queue position in compound semiconductors.

What are the caveats?

The revenue curve mixes product lines. Lumentum's doubling includes telecom and industrial recovery alongside datacenter; the cleanest AI-specific number is Coherent's datacom-and-communications segment (+59%, 79% of revenue). We used company-wide revenue for the ladder and the segment for the causal read, and we label which is which.

Capacity numbers are the sellers'. The 4× Sherman capacity figure, the commissioning timeline for Lumentum's $377.4 million of construction in progress, and "tightness through 2027" come from company disclosures and trade-press reporting of management commentary; no independent audit of wafer capacity exists. The specific InP substrate price points circulating in 2026 press coverage could each be traced to a single originating report, so this piece describes the direction of substrate pricing but repeats no specific wafer price.

The NVIDIA structure's economics are not public. The purchase-commitment volumes, pricing terms and dividend/conversion terms of the preferred stock are undisclosed; the "$2 billion anchors >$5 billion of demand" comparison is ours and uses Lumentum's own public disclosure of the >$5B commitments as the demand side. It is a scale comparison, not a valuation claim.

The geopolitical single point of failure is real but partially hedged. The 70% indium figure is USGS data on refining, not on wafer production — high-quality InP substrate crystallisation is concentrated in Japan (Sumitomo, JX Metal) and the US (AXT, Coherent), and several firms are expanding. The chokepoint is upstream of the chokepoint; that makes it slower to move, not absent.

And the demand curve assumes the optical transition keeps running on schedule. Coherent's CEO ties the growth explicitly to "AI datacenter architectures increasingly transitioning from copper to optical connectivity." If rack-scale copper solutions or co-packaged optics absorb more of the interconnect budget than expected, pluggable laser demand shifts shape — total laser demand does not vanish (CPO uses lasers too, just relocated), but the product mix and margins change. Watch the NPO/CPO commitments, not the transceiver headlines.