The Chip Factory That Hyperscalers Are Quietly—Wait, Openly—Fighting Over

In the fall of 2023, a materials scientist at a mid-tier transceiver company in Shenzhen got a phone call that changed his week. His procurement team had been trying to lock down a supply agreement for indium phosphide wafers (the semiconductor substrate that makes high-speed optical transceivers possible) and had been told, politely but firmly, that the wait time was now measured in quarters, not weeks. He called his counterpart at a competing firm in Taiwan. Same story. Then he called a contact at a hyperscaler's supply chain office in Seattle. The person on the other end laughed. 'We know,' they said. 'We're the ones telling your bosses to secure it.'

This is not the kind of supply chain drama that makes front pages. Silicon gets the headlines. Indium phosphide (InP) does not. But here's the thing: every single optical transceiver inside a modern AI data center relies on InP-based laser chips to convert electrical data into light pulses that travel through fiber optic cables. Without InP, your 800G and 1.6T transceivers do not exist. Your GPU clusters cannot talk to each other at the speeds that matter. Your entire AI training run grinds to a halt.

The physics are stubbornly specific. Silicon is great at logic. It is terrible at generating light. Gallium arsenide works for some applications but falls apart at the wavelengths and speeds that data centers need. Indium phosphide occupies a narrow, essential niche: it can emit, detect, and modulate light at 1310nm and 1550nm wavelengths (the sweet spots for fiber optic communication) while handling data rates above 100 gigabits per second per lane. There is no substitute. Not today, not in five years.

The demand math is staggering. NVIDIA shipped over $60 billion worth of data center GPUs last quarter alone. Each rack of those GPUs needs dozens of optical transceivers to function. The transceiver market, which hovered around $12 billion in 2023, is on track to pass $30 billion by 2027 according to multiple industry forecasts. And every single one of those transceivers starts with an InP wafer.

But here's where it gets interesting. The supply side of this market is extraordinarily concentrated. Making high-quality InP substrates is genuinely hard (we are talking about growing single crystals from a toxic melt at over 1,000 degrees Celsius in a controlled atmosphere, then slicing and polishing them to atomic-level flatness). There are fewer than a handful of companies on the planet that can do this at commercial scale. Most of them are Japanese. One of them is not.

During the first quarter of 2026, something remarkable happened in this tiny corner of the semiconductor supply chain. A company reported that its InP backlog had crossed $100 million for the first time in its history. Its CEO described hyperscalers directly pressuring transceiver makers to lock up supply agreements (not requests, mandates). Demand from China doubled in Q1, and management expected it to double again in Q2, because InP wafers produced inside China for Chinese customers do not require export permits. The company is racing to more than double its production capacity by the end of this year and quadruple it by the end of 2027.

This is a company that has been around since 1986. It was founded by a Chinese-American PhD who spent years perfecting compound semiconductor crystal growth techniques at Bell Labs before going out on his own. For most of its life, it was a specialty materials supplier selling to a small, stable market of telecom equipment makers. The AI boom has fundamentally altered its trajectory.

The financial profile is changing fast. After years of hovering near breakeven, the company guided for GAAP profitability in Q2 2026 and laid out a path to 35%+ gross margins as capacity ramps. Its CFO was appropriately cautious (export permit timing for non-China shipments remains unpredictable), but the CEO's tone was unmistakable: this is the demand wave he built the company to ride.

What makes this story particularly compelling is the top-down pressure dynamic. When Microsoft and Amazon are telling their transceiver suppliers to secure substrate supply, that is not a normal purchase order. That is a strategic directive from companies spending $80 billion or more per year on infrastructure. It means the demand is not speculative. It is being mandated by the largest technology buyers on Earth.

And yet the company producing these wafers trades at a market capitalization that would barely register as a rounding error on NVIDIA's balance sheet.