AI bottleneck goes beyond GPUs
The investable AI story is broadening: analysts now point to networking, bandwidth and other infrastructure as potential choke points rather than only GPUs, and some strategists prefer names outside the pure compute leaders. Hardware vendors are also repositioning — Lenovo is pushing NVIDIA-based systems across edge and data‑center deployments, and Apple has approved eGPU drivers to let AMD/NVIDIA devices run AI workloads on Macs — all signals that the AI stack constraint may be moving away from raw chip supply into interconnects and integration. (247wallst.com, wazzup.ph, tomshardware.com)
The first phase of the AI boom was easy to see. Buy more GPUs. Build bigger clusters. Wait for Nvidia to ship. That story is still real, but it is no longer the whole story. The new constraint is what happens between the chips. Once thousands of accelerators are packed into a training cluster, the problem stops being raw compute and starts becoming movement: how fast data can cross racks, switches, and buildings without stalling the model. That is why the market is suddenly paying attention to optics and networking. A 24/7 Wall St. report published Monday argued that the more interesting AI trade now sits in bandwidth rather than compute, pointing to Lumentum, Coherent, and Fabrinet as the companies selling the lasers, transceivers, and manufacturing capacity needed for the next wave of AI buildouts. The report says Nvidia recently put $4 billion into Lumentum and Coherent, split evenly, and tied those investments to multi-year supply relationships. It also notes a shift toward 1.6-terabit optical links as copper runs into distance and power limits at modern cluster scale. (247wallst.com) That logic matches what the hardware vendors themselves are now saying. Nvidia’s own Spectrum-X Ethernet platform is pitched less as a faster pipe than as a way to keep AI clusters from wasting expensive GPU time. The company says Spectrum-X can improve AI network performance by 1.6 times over standard Ethernet, and its newer cross-data-center version is built to make separate facilities behave like one AI “super-factory.” When the networking vendor starts selling determinism, congestion control, and telemetry as core AI features, the bottleneck has plainly moved. (nvidia.com) Other networking companies are chasing the same opening from different angles. Broadcom’s Tomahawk 6, now shipping, delivers 102.4 terabits per second on a single switch chip and is designed for scale-up and scale-out AI networks, including systems that stretch past a million accelerators. Cisco’s Silicon One G200 makes a similar pitch in plainer enterprise language: 51.2 Tbps, low latency, large buffers, and less packet loss for RDMA and RoCE traffic. These are not side components anymore. They are the machinery that decides whether a giant GPU cluster behaves like one computer or a traffic jam. (broadcom.com) Once that becomes clear, Lenovo’s repositioning makes more sense. In January, the company unveiled its AI Cloud Gigafactory with Nvidia and framed the whole project around speed to deployment and “time to first token,” not just server count. Lenovo said the program combines manufacturing, liquid cooling, storage, ultra-low-latency networking, and services to help cloud providers stand up gigawatt-scale AI factories in weeks. That language matters. It treats AI infrastructure as an integration problem. The hard part is no longer merely obtaining chips. It is turning power, cooling, networking, storage, and software into a machine that produces tokens quickly enough to justify the spend. (news.lenovo.com) The optics suppliers are moving in lockstep. Coherent said in March that it would show 1.6T and 3.2T pluggable transceivers at OFC 2026, plus early work on 12.8T-class architectures beyond that. The point is not just speed for its own sake. It is power-efficient connectivity for AI data centers that are growing too large for older interconnect assumptions. The cluster gets bigger, the links get faster, and the “plumbing” starts to look like the product. (coherent.com) Even Apple’s odd little eGPU story fits this shift. Over the weekend, reports said Apple approved Tiny Corp drivers that let AMD and Nvidia external GPUs run on Apple silicon Macs for AI workloads. The setup is not for gaming, and it is not a polished consumer feature. It is a compute hack that Apple is now willing to sign. The interesting part is not that Macs suddenly became modular workstations. It is that AI demand is pulling hardware companies toward any path that adds usable acceleration, even through awkward external links, as long as the software can make it useful. (appleinsider.com)
