Delos Data's $100M Round Confirms It: The AI Bottleneck Has Moved from the GPU to the Wire

Delos Data builds chips and orchestration software for the cables and switches that tie GPUs together inside AI data centers. CEO Ed Doe told EE Times the company's product, branded Nonstop AI, lets customers run scale-up domains of more than 1,000 GPUs in flexible topologies tuned for inference — and crucially, lets them do it on Ethernet-class hardware rather than NVIDIA's proprietary NVLink/NVSwitch fabric. The platform had its first major showcase at Computex in June 2026, where Delos showed a 36-port OSFP chassis delivering 1.6 Tbps per port.

The capital markets' enthusiasm for an unfunded chip startup is downstream of a re-rating already happening inside public-company earnings. Dell'Oro Group's July 2026 forecast put cumulative AI back-end switch spending at roughly $1 trillion through 2030, with scale-up networking — exactly the domain Delos targets — capturing more than half of that pool. A separate February 2026 Dell'Oro release sized the AI back-end switch market itself at over $100 billion annually by 2030, with Ethernet leading the scale-up segment. Broadcom is already printing that thesis: in its fiscal Q3 2026 (quarter ended Aug 3, 2026), AI semiconductor revenue hit $16.7 billion, up 221% year over year, and management guided Q4 AI semiconductor revenue to ~$21.7 billion (+236% YoY).

The merchant silicon stack for AI back-end switching today is a near-oligopoly. Broadcom sells the Tomahawk family — Tomahawk 6 at 102.4 Tbps for scale-out, Tomahawk Ultra for scale-up, and Jericho-class platforms for scale-across — and Broadcom's silicon shows up in roughly12% of the data-center Ethernet switch market through partner OEMs like Arista Networks. Marvell sells custom XPU ASICs, Ethernet switches, optical DSPs, and co-packaged optics silicon into the same hyperscaler buyers; fiscal2026 data center revenue hit $8.2 billion, 76% of total mix. Astera Labs plays the adjacent layer: signal conditioning (Aries retimers), the new Scorpio X-Series 320-lane fabric switch, and the Taurus Smart Retimer/Redriver family explicitly targeting 224G Ethernet and UALink connectivity.

Delos slots in at the orchestration and reference-design layer — a Broadcom Tomahawk merchant customer, not a competitor to it. But the bigger strategic question is who controls the standards layer. The UALink consortium (members include Astera Labs, Marvell, Broadcom, Intel, AMD, and Cisco) is publishing open scale-up specs to chip away at NVIDIA's NVLink moat. NVIDIA's response: Spectrum-X Ethernet, which went from under 4% of data-center Ethernet switch share two years ago to 21.5% in Q1 2026, and a total networking business (InfiniBand + Spectrum-X) running at ~$60 billion annualized.

AI fabric chips don't ship without the photonics, the DSPs, the SerDes chiplets, and the foundry capacity underneath. The capex reallocation toward interconnect means those tiers get pulled along. Optical DSPs and SerDes: Credo Technology is the pure-play — its Cardinal 1.6T optical DSP on 3nm and active electrical cables (AECs) are what move1.6 Tbps between switches and accelerators; optical revenue is targeted above $600M in fiscal2027. Laser components and transceivers: Lumentum and Coherent supply the lasers and integrated optics that go into transceivers and co-packaged optics modules; Lumentum forecast 65% Ethernet transceiver growth for 2026 in a company blog post. Foundry: TSMC is where most of these chips land — Broadcom's Tomahawk, Marvell's custom ASICs, Astera's Scorpio X, and Credo's DSPs all run on TSMC's advanced nodes.

Three near-term catalysts will tell investors whether the Delos round is a one-off or the leading edge of a capex pivot. (1) Hyperscaler capex prints for Q3 2026: the four largest US cloud providers will report capex over the next four to six weeks; consensus is still anchored to AI training, so any upward revision toward networking and inference infrastructure should re-rate the merchant silicon trio. (2) 1.6 Tbps design wins: Broadcom's Tomahawk 6 (102.4 Tbps) began volume shipments in March 2026; every hyperscaler that places a follow-on order is a datapoint confirming the back-end switch upgrade cycle. (3) InfiniBand-to-Ethernet migration in2026 deployments: Ethernet ports are typically 20–40% cheaper per port than equivalent InfiniBand hardware, and NVIDIA's Spectrum-X demand is reportedly running close to its legacy InfiniBand base for new 2026 builds — that's the green-light signal for UALink and for Astera Labs's Scorpio X-Series.

In the long-term (one to three years), the structural question is whether the interconnect market ends up consolidated around three merchant silicon vendors plus NVIDIA, or whether open-standard UALink breaks the field wide enough for a Delos, an Astera, or a Credo to carve out real share. Dell'Oro's $1T cumulative forecast assumes scale-up alone is more than half of that pool — that's a TAM big enough for multiple winners, but not so wide that fragmentation survives. Watch for Broadcom's networking silicon revenue as a percentage of total AI semi: in Q3 FY2026 it was 27% ($4.5B of the $16.7B AI semi total after subtracting XPU's 73% share); if that share trends toward 35–40% by FY2027, the market is confirming the interconnect is the next silicon battleground.

Delos Data's $100M round is best read as a confirmation, not a catalyst. The catalyst is already in the tape: Broadcom is running a $16.7 billion AI silicon quarter with networking as the second-largest line item, Astera Labs is printing 100%+ growth, and Credo Technology is guiding optical revenue above $600M. The story for the next 12 months is whether the networking slice of AI capex continues to widen. For investors, the cleanest expressions are the merchant silicon vendors that already have scale and design wins — Broadcom and Marvell — plus Astera Labs for the highest-beta exposure to AI fabric switches. For optical, Credo Technology is the pure-play; Lumentum and Coherent are the broader-based picks with bigger legacy revenue bases. Arista Networks is the system-level way to play the same theme. NVIDIA remains the share-gainer of record on Ethernet, and that risk cuts both ways.