by Gaurav Sharma with Alan J Weissberger

Introduction:

Anthropic [1.] earned more revenue in the second quarter of 2026 than it did in all of 2025. The Claude AI maker company raked in $11.5 billion between April and June 2026, up from $787 million in the year-earlier quarter, $4.73 billion in Q1-2026 and up from ~ $10 billion for the entire previous year.  CEO Dario Amodei told CNBC that the growth had been “just crazy” and “too hard to handle,” with demand far outstripping the company’s ability to build infrastructure with demand racing ahead of the company’s ability to scale infrastructure.

Anthropic is not alone. Bank of America projects that AI compute demand will outstrip supply through 2029. GPU lead times now stretch to 36–52 weeks.

Note 1. Anthropic is an American artificial intelligence startup founded by former OpenAI members that focuses on developing safety-oriented, steerable, and interpretable large language models like the Claude AI assistant.

This is happening even as the industry throws historic capital at capacity. Global data-center capital expenditure (capex) is on track to surpass $1 trillion in 2026, according to Dell’Oro Group. Yet the teams building with AI still cannot procure the compute they need, when they need it, at a price that lets them survive long enough to validate their thesis. The AI compute market has a switchboard problem — and fixing it calls for the same kind of thinking that transformed telecommunications.

The Access Gap:

AI-first startups now devote 40 to 50 percent of revenue to GPU hosting and inference compute, and their gross margins sit between 25 and 60 percent — versus 75 to 85 percent for traditional software companies. Compute has become the single largest cost line for most AI businesses, and it dictates what a team can afford to build.

Meanwhile, meaningful enterprise GPU capacity sits dormant. Teams hoard hardware for fear of losing access, locking accelerators into long-term reservations that sit underused overnight and between training runs. The capacity exists; the coordination does not.

The burden lands hardest on those who cannot afford the reservation game. Founders step down to cheaper hardware that slows their research. Teams cut experiments because they cannot secure enough accelerators. Projects stall while usable compute sits idle behind someone else’s contract. In this environment, access — not merit — decides which ideas reach the market and which never get tested.

From Switchboards to Packet Switching:

The early telephone network was run by hand. Every call required an operator to connect the subscriber — and each call claimed a dedicated circuit for its entire duration, even during silence. It worked, but it was slow, labor-intensive and wasteful.

GPU procurement works the same way today. An AI team identifies the hardware it wants, negotiates a reservation with a hyperscaler — AWS, Microsoft Azure or Google Cloud — and waits for capacity to become available. Each commitment locks a slice of the fleet to a single customer. The process is manual, slow and wasteful.

Telecommunications escaped this model in stages. Automated switching removed the operator; packet switching removed the dedicated circuit. Instead of reserving an entire line for one conversation, the network segmented each message into packets and routed them over whatever path had spare capacity, allowing many conversations to share a single trunk through statistical multiplexing. The same physical infrastructure carried far more traffic because capacity was allocated dynamically rather than reserved in advance.

AI compute needs an analogous shift: an orchestration plane that discovers available accelerators across multiple sources and routes each workload to suitable hardware — without the customer negotiating each connection individually.

It is already taking shape. Vendors are building orchestration systems that aggregate capacity from owned infrastructure, data centres and distributed GPU providers, then present it to the customer as a single service. The customer submits a job; the orchestration layer selects suitable hardware, assembles a cluster and delivers it.

Different Workloads, Different Routing:

Orchestration must be workload-aware. Pre-training the largest frontier models requires thousands of accelerators coupled over low-latency fabrics with precise topology; these jobs will continue to demand dense, purpose-built clusters.

Inference, fine-tuning and research are far more elastic. They tolerate geographic spread and run across a wider, more heterogeneous hardware pool. This mirrors how packet-switched networks treat traffic types differently while carrying them on shared infrastructure: a voice call needs bounded latency and continuity, while an email is routed over whatever path has spare capacity.

That differentiation opens the door for network operators. Data-centre operators and carriers already own much of the physical connectivity — fibre, points of presence, interconnection — needed to knit scattered compute into a coherent supply system. Rather than letting AI infrastructure consolidate into a small number of hyperscalers, the industry can use existing transport and edge assets to link regional data centres and GPU providers into a broader, more liquid market.

Robust Data Center Fabric Required:

The trillions in planned capex should be judged by more than the number of GPUs installed. If new capacity flows mainly to customers who can lock in multi-year reservations, the supply of compute grows even as the population of companies able to use it shrinks — fewer experiments, fewer competing hypotheses, a smaller set of teams shaping what AI becomes.

A healthier market would let AI teams reach compute from multiple providers through a single, well-connected service, with the network doing the work of matching each job to the right hardware — the telecoms discipline of statistical multiplexing applied to the GPU fleet.

Telecommunications offers the template. Every forward step it took made the same physical infrastructure serve more users. AI compute looks ready for the same move. The hardware is there, but is the network that connects it robust enough?

The physical network that connects AI compute servers (GPUs/TPUs) to each other and to high-performance storage is collectively called the Data Center Network (DCN) Fabric.  The physical network is split into two primary layers depending on what is being connected:

1. The Backend Network (Compute-to-Compute):

This is the ultra-high-speed, lossless network that connects AI compute servers (or individual GPUs) to one another. It handles “East-West” traffic—such as gradient exchanges and parameter updates—during massive parallel AI training.InfiniBand: Long considered the gold standard for high-performance computing (HPC). It relies on dedicated, high-speed physical switches and host channel adapters (pioneered largely by NVIDIA/Mellanox). It features native Remote Direct Memory Access (RDMA), allowing systems to exchange data directly from memory to memory without involving the host CPU.AI-Optimized Ethernet (RoCEv2): A highly popular open alternative that uses traditional physical Ethernet cabling and switches but runs RDMA over Converged Ethernet (RoCEv2).

Platforms like NVIDIA Spectrum-X utilize optimized Ethernet hardware to achieve lossless, low-latency performance comparable to InfiniBand.Ultra Ethernet: Driven by the Ultra Ethernet Consortium (UEC), this next-generation physical transport standard optimizes Ethernet specifically for massive scale-out AI environments.

2. The Frontend / Storage Network (Compute-to-Storage):

This network connects the AI compute servers to centralized, high-performance storage arrays (like NVMe-oF, SAN, or NAS systems) to stream massive datasets into the GPUs.High-Speed Ethernet: The physical storage network is predominantly built on high-bandwidth Ethernet (moving rapidly up to 400G and 800G per port).Storage Protocols: It leverages protocols like NVMe-oF (NVMe over Fabrics) or RoCEv2 to pull unstructured data (images, text corpuses) from storage units into the compute cluster at lightning speeds without stalling the GPUs.

………………………………………………………………………………………………………………………………………………….

References:

Telecom data centers must be redesigned for the AI era with rack scale architectures, enhanced power & cooling requirements

Analysis: Ethernet gains on InfiniBand in data center connectivity market; White Box/ODM vendors top choice for AI hyperscalers

Will AI clusters be interconnected via Infiniband or Ethernet: NVIDIA doesn’t care, but Broadcom sure does!

Cisco’s Silicon One G300 as the dominant AI networking fabric, competing with Broadcom’s Tomahawk 6 series

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)