AI Data Centers Hit an Interconnect Wall During A Wild July

Photonics have had a wild summer

The constraint on artificial intelligence is no longer the chip. It is the wire between the chips. As AI clusters scale toward hundreds of thousands of GPUs, the industry's attention has shifted from raw compute to the optical links that move data between processors. That shift is redrawing the map of who matters in the AI supply chain.

The transition has not been orderly. A leveraged hedge fund blowup in July dragged the entire photonics complex lower in a matter of days. The selloff hit companies whose demand outlook had not changed. For investors, the episode raised a question that will define the second half of 2026: was the drawdown a warning, or a dislocation?

The Bottleneck Moves to the Network

At $NVDA Nvidia's GTC conference in March, CEO Jensen Huang told the industry that inference demand from hyperscalers is growing exponentially, doubling in roughly a year. Copper cannot carry that load across racks at acceptable power. Industry engineers now argue that every AI data center interconnect will be optical within five years, a view echoed across the OFC 2026 conference floor.

The numbers behind that claim are large. Market researchers value optical interconnect in AI data centers at $9.94 billion in 2025, with projections reaching $41.09 billion by 2035. Co-packaged optics, which moves the optical engine next to the switch silicon itself, is growing faster still. Every major hyperscaler and every major networking chip vendor is now designing around optics as the scale-out layer.

A $45 Billion Unwind as Leopold gets Margin Called

Against that backdrop, July delivered a stress test. Situational Awareness, the AI-focused hedge fund run by 24-year-old former OpenAI researcher Leopold Aschenbrenner, peaked near $45 billion in early July, up roughly 450% for the year on reported leverage as high as 400%. When the AI trade rolled over between July 10 and July 20, margin calls forced the fund to unwind. Its entire public book was sold in a block trade to Ken Griffin's Citadel at a discount, near the lows, according to reporting from CNBC and The Wall Street Journal.

The unwind fed on itself. Names across the AI infrastructure stack fell 30% or more in two weeks. Social media commentary noted that Citadel had publicly flagged the risk of a surprise Fed rate hike days before the collapse, then purchased the very stocks that fear helped crush. That sequence remains market chatter rather than confirmed reporting, and Citadel has not commented. What is confirmed is simpler: a forced seller met a willing buyer, and the tape moved on technicals, not fundamentals.

The Photonics Complex in the Downdraft

Optical names took the beta hard. Coherent $COHR and Lumentum $LITE sold off alongside the compute and power names at the center of the unwind. Smaller photonics companies fell further. Sector coverage of optical chip stocks published mid-month captured the tension: the bandwidth demand story was intact even as the stocks were repriced.

$POET Technologies is a case study in the disconnect. The optical interposer developer declined 29% over the month, in line with the broader AI data center drawdown. The company's operating position had strengthened, not weakened, into the selloff. POET raised more than $400 million in fresh capital this year, struck a supply agreement with Lumilens, and reaffirmed its production ramp for the second half of 2026. Its wafer-scale optical interposer platform targets the exact packaging problem hyperscalers are trying to solve: integrating lasers, modulators, and photonic circuits at a cost and power budget copper cannot match.

The pattern extended across the group. Companies exposed to the optical build-out entered August trading well below their spring highs, while the order books and partnership pipelines behind them kept expanding. Coverage of OFC 2026documented that gap in real time: new transceiver partnerships, record booth commitments, and fresh capital raises landing in the same tape that was marking the sector down.

Dislocation, Not Deterioration

Forced selling answers a liquidity question, not a demand question. Nothing in the July unwind changed the physics of AI clusters or the trajectory Huang described in March. Inference workloads are still compounding. Hyperscaler capital spending on next-generation networks is still rising. The 1.6-terabit transceiver generation is still moving from roadmap to deployment.

The history of technology shakeouts is consistent on this point. Leverage-driven selloffs reset prices faster than they reset fundamentals. The gap between the two is where the next cycle's returns tend to come from. For the optical interconnect layer of the AI build-out, that gap opened in July. The bottleneck did not go anywhere.