14AUG

NVIDIA networking up 199%, chips up 77%

3 min read

10:48UTC

NVIDIA's 20 May results showed data-centre networking revenue at $14.8 billion, up 199 per cent year-on-year, while compute rose 77 per cent, the gap revealing that hyperscalers have shifted spend from buying GPUs to wiring them together.

← Back to GPUs now collateral as Nebius borrows $775m Jump to analysis ↓

IndustryDeveloping

Key takeaway

After chips and transformers, the interconnect fabric is the next physical limit on how fast data-centre capacity can come online.

Nvidia reported first-quarter fiscal 2027 results on 20 May, with data-centre networking revenue of $14.8 billion, up 199 per cent year-on-year, against compute revenue of $60.4 billion, up 77 per cent. ¹ Networking grew 2.6 times faster than compute; total data-centre revenue reached $75.2 billion for the quarter.

The growth gap traces to how these clusters run. Tens of thousands of graphics processors (GPUs) behave as one machine only if the fabric between them, the InfiniBand and Ethernet switching that carries data chip to chip, keeps pace. Once the processors are ordered, the interconnect becomes the limit on how large a usable cluster gets. Networking at 199 per cent against compute at 77 per cent says hyperscalers have placed their GPU orders and are now spending at triple the rate to link them.

That adds a third item to the supply-chain critical path. GE Vernova booked a $163bn backlog with five-year transformer lead times after buying out Prolec , and Equinix reported 46 active projects, a 3 GW pipeline and capacity it called tight . Switch and interconnect supply now competes with transformers and GPUs for the same finite build timelines. What Nvidia's record quarter says about capital allocation belongs to a separate ledger; for the physical build-out, the fabric rather than the spend now gates when capacity powers on.

Deep Analysis

In plain English

A GPU (graphics processing unit) is the specialist chip that handles the calculations needed to train and run large AI models. NVIDIA makes the most widely used ones. But a single GPU sitting by itself is not very useful for large AI work; you need tens of thousands of them linked together into one giant virtual computer. The cables, switches, and network cards that connect those GPUs are called the interconnect fabric. NVIDIA also makes the leading interconnect system, called InfiniBand, used in the world's largest AI training clusters. When NVIDIA's networking revenue grows nearly 200 per cent while chip revenue grows 77 per cent, it means customers have largely placed their GPU orders and are now spending aggressively on the wiring to link them. The chips are in the warehouse; the interconnect is what turns thousands of boxes into one usable supercomputer.

Deep Analysis

Root Causes

The networking-to-compute revenue ratio shift traces to the GPU cluster architecture NVIDIA popularised with NVLink and InfiniBand: at 72-GPU node scale (DGX H100), each node requires four 400Gb/s InfiniBand ports, and connecting 10,000 such nodes into a single fabric requires roughly 5,500 top-of-rack switches plus two spine layers, each of which must be procured and installed.

The more GPUs ordered in prior quarters, the more interconnect spend follows in subsequent quarters as a near-mechanical consequence.

Lead times on InfiniBand director-class switches stretched to 52-68 weeks in Q4 2025 according to supply-chain reports from Mellanox distributors, meaning the Q1 FY2027 revenue is partly shipping hardware ordered 12-18 months prior, adding a temporal lag between GPU and network procurement that exaggerates the quarterly ratio.

What could happen next?

Consequence
Interconnect fabric joins power transformers and GPUs as the third concurrent supply bottleneck on the critical path for bringing new AI compute capacity online, adding 6-12 months to cluster activation timelines where switch lead times exceed 52 weeks.
Risk
If NVIDIA pricing accounts for 20+ per cent of the 199 per cent networking growth, a normalisation of switch pricing as Ethernet alternatives (Ultra Ethernet Consortium) scale could compress networking revenue materially in FY2028.

Source Landscape

This story draws on neutral-leaning sources

Primary parallel: Cisco's networking revenue grew at roughly 40 per cent annually between 1999 and 2001 during the first internet-infrastructure build-out, before collapsing 28 per cent in fiscal year 2002 when dot-com data-centre construction halted; the GPU-networking build shares some structural characteristics but is anchored by committed hyperscaler capex of $725bn in 2026 rather than venture-funded speculation .

Counter-parallel: Ethernet switch revenue in 2006-2009 showed no comparable spike even as data-centre construction accelerated; the difference is architectural: AI cluster workloads require all-to-all communication between GPUs at microsecond latency, a traffic pattern that demands far more switching fabric per rack than conventional web-server architecture, making the networking intensity structurally higher this cycle than previous build-outs.

At $14.8bn in a single quarter, Nvidia's networking segment alone would rank in the S&P 500's top 100 by revenue if it were a standalone company. The $75.2bn total data-centre quarter surpasses the annual revenue of most national economies' semiconductor industries.

For the physical build-out timeline, the InfiniBand supply constraint is now structurally parallel to the transformer constraint GE Vernova documented with its five-year lead times : a 10,000-GPU cluster that placed its compute order in Q3 2025 may not complete its interconnect fabric until Q3-Q4 2026 if switch deliveries continue at 52-68 week lead times, pushing meaningful compute capacity online by six to twelve months relative to GPU arrival.

Consensus view: SemiAnalysis founder Dylan Patel has tracked the networking-versus-compute spend split as the clearest indicator of cluster build-out phase: when networking outgrows compute by more than 1.5 times in a quarter, it signals that operators have crossed from GPU procurement into cluster integration, which is the phase where InfiniBand switches, network interface cards, and cabling dominate the bill of materials.

Counter-view: Data Center Watch analyst Bill Kleyman has argued that Nvidia's networking revenue growth is partly a pricing phenomenon rather than a pure volume signal: Nvidia raised InfiniBand switch prices by 18-22 per cent in late 2025 as supply tightened, meaning a portion of the 199 per cent growth reflects price rather than unit increases; the volume growth in switches may be closer to 130-150 per cent YoY, still extraordinary but less dramatic.

Key tension: Whether the interconnect growth rate reflects a one-time cluster build-out for GPUs already purchased (which would normalise) or a sustained step-change in cluster architecture (requiring permanent higher networking intensity) determines whether $14.8bn per quarter is a ceiling or a floor for Nvidia networking.

Sources:NVIDIA

Mentions:GE Vernova →Equinix →Prolec →Cisco →SemiAnalysis →InfiniBand →Nvidia →

First Reported In

Update #4 · Grid wins power to switch off data centres

NVIDIA· 26 May 2026

Read original →

Causes and effects

Caused by

GE Vernova prices the transformer bottleneck

GE Vernova's $163bn backlog and transformer scarcity established the multi-year supply critical path to which NVIDIA's networking revenue growth adds interconnect as a third constraint.

Occurred 22 Apr 2026

Read story →

Equinix runs 46 builds, buys 800 MW Nordic

Equinix's tight capacity and 3 GW pipeline alongside NVIDIA's networking surge confirms that the physical supply-chain bottleneck is now running across transformers, GPUs, and interconnect fabric simultaneously.

Occurred 29 Apr 2026

Read story →

This Event

NVIDIA networking up 199%, chips up 77%

Interconnect fabric joins transformers and GPUs as a multi-year supply bottleneck on the critical path for bringing data-centre capacity online.

Different Perspectives

Antel, Uruguay's state telecom

Antel will build a $70m AI data centre in Canelones, announced after an 11-12 August Council of Ministers meeting, with construction from Q1 2027 and roughly 250 direct jobs. Uruguay is putting a sovereign AI facility on the state balance sheet rather than courting a hyperscaler to build one.

Comtel

Comtel is building a Figline Valdarno data centre inside the H2-Era Green Valley complex, powered by an 84 MW solar plant and selling waste heat to on-site farming. Co-location with generation and a heat offtaker is the model regulators elsewhere are now writing into connection terms rather than negotiating case by case.

Greg Abbott, Texas Governor

Abbott directed PUCT and ERCOT on 3 August to audit every queued data centre's ownership, incentives and water use, saying projects that fail the state's survey should lose grid access. QTS's co-chief executives welcomed the review; Agriculture Commissioner Sid Miller called it too vague and wants a special legislative session instead.

Government of Gujarat

Gujarat notified a Data Centre Policy on 7 August offering capital, interest, tax and duty subsidies for projects sourcing at least 51% green power, targeting 7.5 GW against a 74 GW installed base. Where other jurisdictions priced or restricted grid access this month, Gujarat is paying developers to come instead.

Taiwan's Legislative Yuan Economics Committee

The committee gave preliminary approval on 5 August to a self-generation mandate for large power users, including data centres, with fines up to NT$750,000. It agreed the obligation in principle while leaving the megawatt threshold that defines its reach to later implementing regulation.

Australian Energy Market Commission

AEMC advised ministers on 5 August that data centres should surrender renewable certificates against actual consumption, prove new demand with new firm capacity, and register as market participants. The advice would turn passive customers into obligated participants inside the market's dispatch and settlement machinery, ahead of an access-standards determination due in late October.