Hurlston Just Validated the Thesis on Bloomberg
Wow this was a very busy week of AI announcements. This morning in Tokyo, Lumentum CEO Michael Hurlston sat down with Bloomberg and said something that should not be possible to say as the CEO of a publicly traded company: “The capex numbers from the US hyperscalers are enormous and there seems to be no end in sight. We’re falling further and further behind the demand. We would be sold out though all of 2028 within two quarters.” Read that last sentence again. Within two quarters — by roughly Q2 or Q3 of this year — Lumentum’s entire 2028 order book will be filled. The company is repurposing old electronics fabrication facilities to bring new capacity online in half the time it would take to build from scratch. Japan capacity expansion is committed at $100 million and likely scaling to $250 million.
Lumentum CEO Michael Hurlston on Bloomberg from Tokyo, April 10, 2026. Full interview at Bloomberg.
This comes on top of what Hurlston had already disclosed on the Q2 earnings call: an incremental multi-hundred-million-dollar CPO order deliverable in 1H27, an OCS backlog exceeding $400 million, and both product lines described as “only at the starting line.” Combine that with NVIDIA’s $2 billion investment, the multi-billion purchase commitment, and Lumentum’s March 23 S&P 500 inclusion, and the structural picture is clearer than it was three weeks ago at OFC. The thesis hasn’t changed. It just got stronger in public, from the CEO of my largest optical position, on live television — and it is exactly what I laid out in The Token Dollar: scaling laws are holding, demand curves are not bending, and the dollar-denominated supply chain sits at the center of it.
The NeoCloud Bear Case Just Got Resolved
One more same-morning data point, because it tests one of my own prior bear cases. A month ago in The NeoCloud Hypothesis, I flagged a specific risk to NVIDIA-anchored neoclouds like CoreWeave: Claude migrating entirely onto Google TPUs and Amazon Trainium. Anthropic had just signed a $30 billion TPU deal and was ramping Trainium in parallel. If the frontier lab with the fastest-growing workload in AI decided it didn’t need NVIDIA for production inference, the neocloud thesis took a real hit. Today, Anthropic signed a multi-year deal with CoreWeave — a pure-play NVIDIA shop with no TPU, no Trainium, no custom silicon hedge anywhere in the relationship.
The takeaway: even the labs betting most aggressively on custom silicon still need the fastest NVIDIA deployment layer for the near-term workload. The bear case — Claude migrating entirely off NVIDIA — just got directly falsified by the company that would have had to execute it. And every fiber link, every laser, every transceiver inside that CoreWeave footprint is now in higher demand. The neoclouds buy optics. The hyperscalers buy optics. Hurlston’s order book has another customer vector pointing at it, from a different angle, confirming the same structural conclusion he described on Bloomberg this morning.
Beyond the Public Names: Where I’m Looking Privately
One more note before the paywall. The optical thesis doesn’t stop at LITE, COHR, CRDO, and TSEM. I’m actively tracking four private companies attacking bottlenecks from outside the public market — Ayar Labs, Litrinium, Arago, and Avicena Tech — and meeting with management as primary research. The full deep dive on each is below. This is the kind of primary research that doesn’t make it into most semiconductor newsletters, and it’s one of the reasons BEP Research exists.
For paid subscribers: the Mythos 7x compute math and what it means for optics, the full Lumentum thesis, Tower’s manufacturing dominance, CPO reliability data, my Credo reassessment, the private market deep dive, and the complete Optical Cold War analysis with NVIDIA’s $6B bingo card vs. Broadcom’s $200B Google LTA.
Leave a Reply