We’re back! (Indeed, we never left). Nvidia (NVDA:NASDAQ) and Elastic (ESTC:NASDAQ) earnings trumped my 🇦🇺 vacation, for now.

A short holiday week is expected to be, strangely, jam packed with news and earnings.We provide this note with the hope that it’ll help journalist now what’s coming and, with high hopes, what it all means. The full archive and research database live here: https://epistrophy.beehiiv.com (your email address works as a password). Check it out.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

1. Companies Discussed

2. Nvidia’s Networking Nova

   1. The Merge That Moved the Needle

   2. Why Now?

   3. The Real Traffic Winners

   4. What’s Next

3. Tweet O’ The Week

4. Epistrophy In The News

5. 📆 of Epistrophy Events

6. Availability This Week

Nvidia’s Networking Nova

How $8.2B In Networking Sales Predicts The Future

The most important number in Nvidia’s latest earnings wasn’t about Blackwell, generative AI or the shape of next year’s GPU roadmap. It wasn’t even about compute. The quarter’s real tell was a line item that almost no one on the Street led with: networking. Interconnect, optics and switching are now Nvidia’s fastest-growing businesses, and the surest indicator of where AI infrastructure spending is heading. The last decade of AI was driven by compute. The next decade will rise or fall on data movement.

Nvidia reported $8.2 billion in networking revenue, up 162 percent from a year earlier, driven by InfiniBand, 800G optical modules, Spectrum-X Ethernet and NVLink systems. That number makes Nvidia the world’s largest data-movement company. It now generates more networking revenue than Cisco (CSCO:NASDAQ), Arista (ANET:NYSE), Ericsson (ERIC:NASDAQ) or Nokia (NOK:NYSE).

“Our networking business, purpose-built for AI and now the largest in the world,” Huang said in the Q3 earnings conference call. “We are winning in data center networking as the majority of AI deployments now include our switches, with Ethernet GPU attach rates roughly on par with InfiniBand.”

The message is simple. Compute remains essential, but the constraints that shape AI performance now sit in the fabric — the optical and electrical pathways that bind GPUs into something resembling a single machine. Through the lens of last quarter’s networking numbers, his report will look at Nvidia's business to figure out the next big AI winners.

The Merge That Moved the Needle

For most of its history Nvidia was a chip company. Systems were accessories to silicon. That changed in April 2020 with Nvidia’s $6.9 billion acquisition of Mellanox after a bidding war with Intel (INTC:NASDAQ).

Mellanox brought InfiniBand, RDMA, adaptive routing, congestion control, FPGA-grade packet engines and switching silicon built for deterministic performance at scale. It also brought a worldview: that compute is useless if the data cannot arrive predictably. Large-scale AI has now validated that idea completely.

More than technology, the acquisition gave Nvidia control of the one part of the system that scales with model size. GPUs doubled and redoubled in performance. Model sizes leapt from billions to trillions of parameters. Training spilled beyond a single node to pods, then racks, then rows. Each of those shifts required an interconnect capable of keeping pace.

The deal turned Nvidia’s product line into a full communication stack. InfiniBand supplies lossless transport, strict congestion control and packet-level routing. Spectrum-X brings RDMA over Converged Ethernet, priority flow control and hardware-based telemetry to Ethernet clusters. NVLink and NVSwitch bind dozens of GPUs into memory-coherent domains. Together they form a platform in which compute is only half the machine. The other half is movement.

The acquisition has since proved to be one of the most prescient in the history of semiconductor M&A.

Why Now?

Two years ago copper still carried a significant share of short-reach traffic inside AI clusters. Retimers and equalizers helped the signal stretch far enough for dense racks. That window has closed. Blackwell pushed GPU performance so far ahead of the fabric that copper and conventional Ethernet could no longer deliver data at the rate the chips consumed it.

Nvidia’s released a white paper in July that laid out the issue. “Networking for the Era of AI,” puts it plainly: “The network defines the data center.”

It illustrates several constraints converging to force the transition:

• AI is now a distributed-computing problem.
  A single training job spans hundreds or thousands of GPUs. Every step requires collective operations — all-reduce, all-gather, broadcast — that must complete before the next iteration begins. One slow link becomes the system’s runtime.

• Dropped packets stall entire clusters.
  In AI, when a packet drops, it must be retransmitted. Because GPUs synchronize at each iteration, a single dropout forces every GPU in the cluster to wait. Performance collapses.

• Ethernet’s default behavior is lossy.
  Standard Ethernet drops packets when buffers overflow. AI workloads cannot tolerate this. Spectrum-X extends lossless behavior through RDMA — which bypasses the CPU and cuts latency — and priority flow control, which pauses sender NICs before congestion builds.

• Deep-buffer switches increase tail latency.
  Telecom networks use large buffers to absorb bursts. AI workloads need predictability. Deep buffers add jitter that becomes the ceiling on scalability. Shallow buffers and cut-through switching are required.

• Cut-through switching demands uniform speeds.
  If host-to-leaf links run at different speeds than leaf-to-spine links, the switch cannot cut through. It must buffer and retransmit the packet. AI clusters cannot absorb that delay.

• Elephant flows saturate links.
  AI workloads move large tensors — sometimes gigabytes per iteration. These elephant flows consume the full bandwidth of a link. Without intelligent routing they collide and create congestion spikes.

• Flow-level hashing fails.
  Traditional routing hashes an entire flow to one path. If that path is congested, latency spikes. Nvidia’s fabrics route at the packet level, scattering traffic across multiple paths and reassembling it at the destination. This avoids hot spots.

• ECN reacts too slowly.
  Explicit Congestion Notification only signals after congestion appears. AI workloads build congestion too quickly for ECN to respond.

• Telemetry-driven NIC control is required.
  Nvidia pushes switch telemetry directly into the Network Interface Card’s injection logic. The NIC slows transmit rates before congestion occurs. This prevents packet drops at the source.

These constraints all point to the same conclusion: large-scale AI requires deterministic, lossless fabrics. Copper cannot provide them. Commodity Ethernet cannot provide them without modification. Only integrated fabrics — like InfiniBand and the tuned version of Ethernet inside Spectrum-X — meet the requirement.

The industry crossed that threshold this year. Nvidia happened to own the stack built for it.

The Real Traffic Winners

The shift to optical, lossless fabrics does not lift all vendors equally. It rewards the companies whose products sit directly in the signal path.

Fabrinet (FN:NYSE) sits in the middle of the optical module boom. It assembles a large share of the 400G and 800G optics used in AI clusters. Nvidia’s demand is steady and high value.

Coherent (COHR:NYSE) and Lumentum (LITE:NASDAQ) both benefit as 800G optics move toward tighter tolerances, higher power budgets and more complex photonic packaging. They supply the lasers and modulators inside the modules. Yield determines margin, but demand is consistent.

Corning (GLW:NYSE) gains on two fronts. Inside the data center, single-mode fiber is replacing copper for high-density short-reach runs. Outside, AI campuses are being connected to each other with new metro-scale fiber lines. Corning serves both markets.

Broadcom (AVGO:NASDAQ) remains central to Ethernet inference clusters. Its SerDes roadmap defines the lane speeds that enable 800G and, soon, 1.6T modules. Marvell (MRVL:NASDAQ) participates through PAM4 DSPs and cloud-tuned switch silicon.

Jabil (JBL:NYSE) benefits as optical modules and switch systems move toward mid-board optics and tighter thermal constraints. Its experience in precision optical assembly matters.

Applied Optoelectronics (AAOI:NASDAQ) holds a unique position. Amazon’s investment and supply agreement guarantee multi-year volumes for certain optical SKUs. Pricing remains competitive, but demand is locked in.

Legacy networking vendors have opportunities, but not yet the products needed inside AI data centers.

Arista (ANET:NYSE) dominates Ethernet in cloud inference, but the training clusters that consume the most optics remain outside its reach.

Cisco remains the leader in enterprise networking, but does not yet offer the lossless fabrics, packet-level routing or congestion-aware NIC integration required for large-scale AI.

Ciena (CIEN:NYSE) and Nokia (NOK:NASDAQ) build world-class long-haul coherent optics. These systems connect data centers to one another, not GPUs to GPUs. Both have credible pathways into short-reach optical gear, but the shift has only begun.

These companies are not shut out. They are simply not yet present inside the new data-center architecture. The opportunity is real, but so is the gap.

What’s Next

Three outcomes are already visible.

1. AI networking will grow 40–55% annually through 2027.
Model sizes keep rising. Cluster sizes keep expanding. Optical module shipments will likely double again in 2026 and may double once more in 2027 as early 1.6-terabit modules appear. NVLink 5’s move to optical signaling will add additional demand.

2. Nvidia will hold more than 75 % share of AI fabrics inside the data center — but its networking strength ends at the data-center door.
InfiniBand and NVLink remain unmatched for training. Spectrum-X gives Nvidia a presence in cloud Ethernet. But Nvidia does not compete in long-haul transport or metro interconnect. It owns the cluster, not the network between clusters.

3. The supplier winners are clear.
Fabrinet, Coherent, Lumentum, Corning, Broadcom, Marvell and Jabil sell the components that create bandwidth. Their fortunes rise with traffic. Optical tolerances are tightening. Module counts are rising. The value sits in the signal path.

Scaling AI models now depends as much on data movement as on compute. The next decade of AI will be shaped not by bigger chips but by faster links — by the fibers, lasers and fabrics that let thousands of GPUs behave like one.

Nvidia saw that shift coming. This quarter made it visible to everyone else.

Tweet O’ The Week

Epistrophy In The News

I joined NewsNation twice this week, where we discussed Nvidia’s earnings and the future of AI as well as Elon Musk’s latest kooky predictions about humanoid robots.

On Yahoo Finance TV, I joined Josh Lipton to talk about earnings from Nvidia and the how this different this AI bubble is from the accounting frauds that typified the Dot Com bubble. It’s a topic I expect we’ll be visiting quite a bit over the next few years.

📆 of Epistrophy Events

Availability This Week

I’m in San Francisco or down in Silicon Valley all week. Hit me up by text for time-sensitive stuff; email works too, replies later in the day.

Written reports are available to clients, with video summaries on YouTube, and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.