GTC Taipei 2026 | Jensen Huang Took Computing Apart: In the Agent Era, Optical Interconnect Is the Critical Path to Revenue
The star of GTC Taipei 2026 wasn't Vera Rubin. It was Jensen Huang's new definition of "computing": agentic AI is essentially a computer taken completely apart and scattered across every corner of the data center. Once compute is disaggregated, the interconnect layer that stitches it back together (NVLink, Vera CPU, CPO photonic switches) turns from a cost item into the critical path that determines revenue. For Taiwan's optical communications supply chain, the real signal of this keynote hides in a line he repeated more than a dozen times: "every watt is revenue."
1. The Real Shift Isn't the GPU, It's a Redefinition
It's easy to get buried in product count during this keynote: Vera Rubin, Vera CPU, Spectrum-X Photonics, Nemotron, Cosmos, RTX Spark, humanoid robots — six or seven product lines in one talk. But if all you remember is "NVIDIA launched a bunch of stuff again," you missed the thread that ties everything together.
Jensen spent most of the keynote doing one thing: redefining what "a computer" is. Applications used to be binaries running on an operating system; today's application is called an agent, made of four things — a large language model (the thinking), a harness (the coordination, like a body), tools (spreadsheets, browsers, databases, CUDA-X libraries), and a runtime.
The key isn't these four pieces themselves, but where they run.
An agent is not a program executing on a single machine. It is disaggregated, distributed compute — thinking happens on this machine, tool calls on that one, memory on a third.
To a GPU vendor this is a technical description; to the optical supply chain it is an order forecast. Because "disaggregated + distributed" boils down to one conclusion: machine-to-machine connections are about to explode.
2. The Cost of Disaggregated Compute Becomes the Interconnect Bill
Spread a computer across a data center and you gain flexibility — but the price is that data has to keep moving between the parts. Jensen was blunt in the Vera CPU segment: when compute is disaggregated, the network becomes the bottleneck; bandwidth and latency between CPU cores, between CPU and storage, and between CPU and GPU directly decide how fast the whole agent runs.
He summed up the CPU's new role in one line: "agents are impatient." CPUs used to be designed for humans, who think in seconds; agents wait in nanoseconds, and every extra moment stalls the next step. So Vera CPU isn't designed to cram in more cores for rent, but to push single-thread latency to the limit — claiming the world's highest IPC (decoding and executing 10 instructions per clock), 1.2 TB/s of LPDDR5x bandwidth, PCIe Gen6, and 88 Olympus cores linked by a second-generation coherent fabric, with core-to-core bandwidth reaching 3.6 TB/s.
The point here isn't the spec sheet, but the shift in design philosophy: from maximizing throughput to minimizing latency, from renting to humans to feeding agents. SQL 3x faster, NYSE real-time streaming 6x faster, agentic sandboxes 1.8x faster than x86 — these numbers all point to the same thing: the speed of moving data is now money.
And when moving data becomes the most expensive, most critical part of the system, copper can't keep up. Light has to come in.
3. CPO Is No Longer a Demo — It Shows Up in the Network Layer of Production Racks
While breaking down the seven new Vera Rubin chips, Jensen called out one thing: NVIDIA Spectrum-X Ethernet Photonics, which he called "the world's first CPO (co-packaged optics) Ethernet switch". It is paired with ultra-high-power laser modules on indium phosphide (InP), chip-scale packaging, and TSMC's CPO COUPE.
Put these terms together and the signal is clear. The biggest doubt about CPO in recent years has been "crying wolf" — plenty of demos, volume production nowhere in sight. But this time it doesn't sit in some box on a roadmap; it is listed as one of the network-layer components of a Vera Rubin rack already in volume production. CPO has never been won or lost on whether the optics can be made, but on whether packaging can integrate the PIC and EIC at high enough yield and seat the laser reliably next to the switch — we broke this down fully in CPO Is Won in Packaging, Not in Optics.
The laser choice is worth noting. High-power laser modules on InP echo what we observed earlier in The CPO Commercial Year Officially Kicks Off — TSMC COUPE Volume Production and the 200G EML Bottleneck: the light source (laser / EML) has always been the most invisible gating item for CPO volume production. By bundling InP lasers, chip-scale packaging and TSMC co-processing into one production package, NVIDIA is effectively endorsing the entire silicon photonics packaging supply chain — some players on this chain will win real orders, others will be left stranded at the demo stage.
4. Feeling the Speed and Scale of This Chain in Numbers
The keynote offered a few hard numbers:
Assembly time 2 hours → 5 minutes: A Grace Blackwell rack took two hours to assemble; Vera Rubin, by switching to a PCB midplane and removing most cables, now takes five minutes. This isn't just about speed — it makes "cableless" design the core of reliability and serviceability.
Supply chain scale ×2: Jensen said the supply chain built for Vera Rubin is twice the size of Grace Blackwell's, with 150 supply chain partners across Taiwan and millions of square feet of factory space already online.
A single GW AI factory = US$50–100 billion: and this capex "must succeed the first time."
Every watt is revenue: This was the line of the keynote. Jensen stressed repeatedly that a data center's power ceiling is fixed — 1 GW is 1 GW — so how many tokens each watt produces directly equals revenue. Choosing the wrong architecture just because the chips are cheap won't pay off in the end.
Stack the last two points together and you see why optics matters so much. Under a fixed power budget, copper's power consumption and signal loss eat watts that could otherwise compute tokens; CPO brings light right next to the switch chip, and every watt saved can be converted into billable compute. Under the "every watt is revenue" framework, CPO isn't showing off — it takes power budget back from data movement and hands it to the compute that makes money.
5. A Dose of Cold Water: Beyond the Hype, Three Unaddressed Challenges
Three things in this keynote deserve some reservation:
First, CPO yield and serviceability have not been proven in the field. "World's first" sounds great, but CPO's biggest pain point has always been how to replace a failed optical component — co-packaging means the optics are bonded to the ASIC, making field repair far more troublesome than pluggable optical modules. A production announcement ≠ reliable operation at scale, and Jensen himself admitted in the reliability segment that it is "very hard."
Second, token economics may repeat the "inference is easy" underestimate. He recalled that in the Grace Blackwell era everyone said inference was easy, only to find that achieving high responsiveness and high throughput at the same time is extremely hard. Now he describes agentic "disaggregated inference" as if it were solved, but agents chaining multiple tools, running for long periods and shuffling memory in and out will only make overall reliability harder, not easier.
Third, this is a home-turf COMPUTEX keynote, so the narrative naturally leans bullish and fast. Wording like "already in full production" still faces a time gap — yield ramp, qualification, customer adoption — before it lands as an actual shipment curve for the supply chain.
6. What It Actually Means for Taiwan's Supply Chain
Filter out the noise, and the keynote's signal for Taiwan's optical communications supply chain condenses into three points:
The interconnect layer goes from supporting role to lead. When compute is disaggregated, the value weighting of technologies that "stitch the parts back together" — NVLink, CPO, Spectrum-X — jumps. Those making optics, packaging and high-speed connectivity are in the right position.
The battle for CPO supply chain positions has officially begun. InP lasers, silicon photonics PICs, advanced packaging, TSMC co-processing — NVIDIA has already drawn the combination it wants. Companies that make it into this BOM win real orders; those that don't stay in the demo pool.
"Every watt is revenue" will in turn force the pace of optical penetration. As long as the AI factory's bottleneck is power, every step of optics replacing copper has a clear financial rationale — which pushes CPO forward more than any technical argument.
What this keynote truly declared isn't that NVIDIA got several times faster, but this: the moment compute was taken apart, optical communications was promoted from the data center's plumbing to the main structural engineering that determines profit. For STT readers, what to watch next isn't GPU benchmarks, but which Taiwanese suppliers actually landed part numbers on that "world's first" CPO switch.
This article is for technology and industry trend analysis only and does not constitute investment advice.
Comments