Dell’Oro Forecasts AI Back-End Switch Spending to Surpass $100 Billion by 2030

Dell’Oro Group appeared in a Google News headline claiming AI back-end switch spending would approach $1 trillion by 2030. The available forecast says something materially different. Dell’Oro expects spending on switches for AI back-end networks to surpass $100 billion by 2030, not $1 trillion.

The larger figure belongs to a broader category. Dell’Oro has separately projected worldwide data center capital expenditure at $1.7 trillion in 2030. That total covers far more than network switches. It includes accelerated servers, conventional servers, storage, networking, and physical infrastructure.

This distinction matters because a tenfold error changes the story. The verified forecast still describes a large market, but it does not put network switches near trillion-dollar scale. The important contest is therefore not whether switches consume almost $1 trillion. It is how Ethernet vendors are taking a growing share of the networking budget once dominated by Nvidia’s InfiniBand.

What the Google News Headline Gets Wrong

The headline appears to combine a switch-market forecast with Dell’Oro’s much larger data center capital-expenditure outlook.

The Google News result attributes a near-$1 trillion forecast to spending on the AI back-end switch market. However, reporting based on Dell’Oro’s research places the corresponding switch figure above $100 billion by 2030. That is still a major expansion, but it is one-tenth of the headline claim.

The wording also leaves an important measurement question unresolved. A forecast can describe annual spending in 2030 or cumulative spending across several years. Those are not interchangeable measures. Published summaries should state the period clearly before readers compare the number with other infrastructure markets.

Dell’Oro’s broader forecast provides a likely source for the confusion. The firm expects worldwide data center capital expenditure to reach $1.7 trillion in 2030, according to coverage of its five-year outlook. Accelerated servers could represent as much as two-thirds of that infrastructure spending by the end of the period.

That broad capital-expenditure category includes the processors, memory, storage, electrical systems, cooling equipment, buildings, and networks needed to operate data centers. A back-end switch is only one component within that stack. Treating the categories as equivalent greatly overstates the addressable switch market.

The most defensible reading is straightforward. Dell’Oro expects AI networking to become a much larger part of data center spending, while total data center investment approaches the trillion-dollar range. The firm does not appear to predict that back-end switches alone will approach $1 trillion.

A second number supports that interpretation. Earlier Dell’Oro analysis projected more than $100 billion of investment in AI back-end networks over a five-year period. It also expected total data center switch sales, including traditional front-end networks, to exceed $180 billion during that period.

Front-end networks connect general-purpose servers, storage, management systems, and users. Back-end networks link accelerators during training and inference. The two networks serve different traffic patterns, even when both use Ethernet.

The difference is especially important for investors and infrastructure buyers. A $100 billion opportunity can support meaningful growth for switch silicon, optics, network interface cards, and systems. A $1 trillion switch market would imply an entirely different allocation of data center capital.

Readers should therefore treat the syndicated Google News title as an inaccurate summary, not as a verified statement from Dell’Oro. The corrected number does not eliminate the underlying story. It makes the story more credible and more useful.

That story begins with an architectural shift. AI clusters require networks that keep thousands of accelerators synchronized. As those clusters expand, the network increasingly determines how efficiently operators use their most expensive computing equipment.

Why AI Back-End Networks Are Attracting So Much Spending

AI networking is growing because idle accelerators turn network delays into visible financial losses.

A back-end network is the high-bandwidth fabric that carries data among GPUs and other accelerators working on the same job. It differs from the front-end network that connects applications, users, storage services, and conventional servers.

Large AI workloads divide computation across many accelerators. Those processors repeatedly calculate results, exchange data, and wait for the next computation phase. A delayed message can hold up every accelerator participating in the operation.

This behavior makes tail latency important. Tail latency measures the slowest communications in a group, not the average transfer. If one path becomes congested, the entire distributed job can pause while expensive processors wait.

Network utilization also has a direct effect on job completion time. A faster fabric can reduce idle periods and finish a training run sooner. A poorly configured fabric can waste accelerator capacity even when the cluster contains enough nominal computing power.

Scale-out expansion is one driver behind Dell’Oro’s forecast. Scale-out networks connect more servers and accelerator racks into a larger cluster. They require additional switches, network interface cards, cables, and optical transceivers as the cluster grows.

Scale-up designs create another source of spending. These connections allow accelerators within a rack or tightly integrated system to behave like a larger computing unit. Nvidia’s NVLink is the best-known proprietary example, while UALink and Ethernet-based initiatives are pursuing alternatives.

Dell’Oro also identifies scale-across architectures as a future growth area. Scale-across networking links separate data centers so operators can coordinate workloads across sites. The approach responds to limits involving land, electrical power, and the size of individual facilities.

Every expansion adds technical pressure. Longer distances require more optical equipment. Larger clusters create more possible congestion points. Higher link speeds increase power, cooling, testing, and signal-integrity demands.

Port speeds are consequently moving faster in AI fabrics than in conventional enterprise networks. Dell’Oro data presented in a 2026 networking roadmap indicates that most AI back-end Ethernet ports should reach 1.6 terabits per second during 2027. The roadmap extends to 3.2-terabit connections by 2030.

These speeds do not mean every enterprise will deploy a massive AI cluster. Hyperscalers, model developers, sovereign AI programs, and specialized cloud providers account for much of the demand. Many ordinary businesses will consume that infrastructure through cloud services.

The forecast is still relevant to enterprise buyers. Network design affects the availability, performance, and cost of the AI services they purchase. It also shapes whether smaller operators can build competitive clusters without adopting a single vendor’s complete stack.

The broader spending cycle reinforces that pressure. Dell’Oro expects worldwide data center capital expenditure to reach $1.7 trillion in 2030, as summarized in the firm’s data center outlook. Hyperscalers remain central, but specialized cloud providers and sovereign projects add new demand.

That figure offers the right context for the switch forecast. Networking is a critical part of a much larger infrastructure buildout. It is not the entire buildout.

Ethernet Has Become the Main Challenger to InfiniBand

The central market conflict is Ethernet’s expanding ecosystem against InfiniBand’s established performance and Nvidia integration.

When Dell’Oro began tracking AI back-end networks in late 2023, InfiniBand reportedly held more than 80 percent of the market. InfiniBand had a long history in high-performance computing and offered low latency, high throughput, and mature support for collective operations.

Nvidia strengthened that position through its acquisition of Mellanox. The company could sell accelerators, network interface cards, switches, cables, and software as a coordinated system. Customers building large clusters had a proven path with fewer integration decisions.

Ethernet entered the contest with different advantages. Operators already understood the technology, multiple vendors supplied compatible equipment, and existing tools supported its operation. Ethernet also reduced dependence on a single networking architecture.

Those benefits were not enough on their own. Traditional Ethernet could suffer from packet loss, congestion, and unpredictable latency under synchronized AI traffic. Large training jobs expose those weaknesses because many accelerators communicate simultaneously.

Vendors responded by changing the transport, congestion-control, telemetry, and load-balancing layers. Remote direct memory access over converged Ethernet, commonly called RoCE, lets systems move data between memory locations without relying heavily on host processors.

RoCE improved performance but demanded careful configuration. Priority Flow Control can pause selected traffic classes to prevent packet loss. Poor tuning can spread congestion or create head-of-line blocking, where unrelated traffic waits behind a stalled flow.

The Ultra Ethernet Consortium is developing a more coordinated answer. Its members include major chip designers, cloud operators, networking suppliers, and system manufacturers. The group released the initial Ultra Ethernet specification in June 2025.

The specification defines a communications stack for AI and high-performance computing. It covers transport behavior, congestion control, security, software interfaces, and optional in-network collective operations. Those collective functions let the network assist with calculations shared across many accelerators.

Ultra Ethernet Transport also supports multipath packet spraying. Instead of assigning an entire data flow to one route, the system can distribute packets across available paths. This design seeks to improve utilization and avoid individual congestion points.

Flexible ordering is another important feature. Conventional protocols often insist that every packet arrive in a strict sequence. AI workloads can sometimes tolerate different ordering rules, which give the network more freedom to route traffic efficiently.

The specification is not proof that every implementation will match InfiniBand. Standards describe behavior and interfaces, while production performance depends on silicon, firmware, topology, software, and operational discipline. Interoperability also requires testing across products from different vendors.

Dell’Oro nevertheless sees Ethernet as the long-term winner in AI back-end networking. Reporting on the forecast says Ethernet overtook InfiniBand in 2025. That would represent a rapid reversal from InfiniBand’s dominant position in late 2023.

The shift does not necessarily exclude Nvidia. Nvidia sells both Quantum InfiniBand and Spectrum-X Ethernet. It can participate in customer movement toward Ethernet while defending its integrated accelerator and networking platform.

Spectrum-X combines Nvidia Ethernet switches, network adapters, software, and telemetry. Nvidia says the system improves performance over standard Ethernet, but its comparisons are vendor claims and depend on workload and configuration.

The company is also adding co-packaged optics, which place optical components close to the switch silicon. This design reduces the electrical distance between the switch chip and optical connection. Nvidia plans both Ethernet and InfiniBand products using the technology.

According to Nvidia’s photonics announcement, Spectrum-X configurations will offer up to 400 terabits per second of total throughput. The company has described 2026 availability for its Ethernet photonics switches.

Arista, Broadcom, Cisco, AMD, Marvell, HPE, and other suppliers also have strategic interests in Ethernet’s expansion. Some sell complete switches, while others provide switch chips, network adapters, optics, or software.

Ethernet’s success therefore broadens the potential supplier pool. However, it does not guarantee that spending becomes evenly distributed. Large operators can still prefer tightly integrated systems, and vendors can differentiate above an open link standard.

What the $100 Billion Forecast Does Not Guarantee

A growing market does not settle the questions of performance, supplier concentration, power consumption, or investment returns.

Forecasts depend on assumptions about accelerator deployments, cluster sizes, port speeds, and equipment replacement cycles. Each assumption can change before 2030. A slowdown in data center construction would reduce networking demand even if networks retained their share of each project.

The largest uncertainty involves AI economics. Operators are spending heavily because they expect demand for training and inference to justify the infrastructure. Weak utilization, slower revenue growth, or efficiency gains could alter their construction schedules.

Improved models can push demand in both directions. Greater efficiency might reduce the computing needed for a fixed task. Lower operating costs can also encourage more usage, creating new demand through a rebound effect.

Networking requirements also vary by workload. Training a large model creates synchronized communications across many accelerators. Inference can involve smaller groups, different latency targets, and more geographically distributed demand.

Dell’Oro expects training to account for a smaller portion of accelerator resources over time as inference expands. That shift can change topology choices without eliminating network growth. Buyers may emphasize predictable latency, multi-tenancy, and geographic connections rather than building only enormous training clusters.

Ethernet still faces a technical burden. AI traffic often arrives in bursts, and congestion can reduce the performance of the entire job. A 2026 research paper examining modern high-performance fabrics identified congestion as a significant limitation across heterogeneous workloads.

The challenge becomes harder at scale. A configuration that performs well across hundreds of accelerators does not automatically retain the same efficiency across tens of thousands. Failures, uneven links, and traffic collisions become more common as systems grow.

Open standards also need interoperable products. Ultra Ethernet has a published specification, but buyers still require compatible adapters, switches, cables, software, and management tools. They need evidence that combinations from different suppliers operate reliably.

The consortium’s specification history shows continued revisions. Version 1.0.3 arrived in July 2026 after earlier corrections and clarifications. That maintenance is normal for a technical standard, but it also shows that implementation work continues.

InfiniBand remains a credible option during this transition. It has mature deployments, established software integration, and purpose-built features for high-performance computing. Buyers running demanding training jobs can prioritize predictable performance over supplier diversity.

Nvidia’s strategy reduces the contest to something more complex than Ethernet versus Nvidia. The company supports Ethernet through Spectrum-X while continuing to develop InfiniBand. It can use common software and system relationships across both product families.

Scale-up networking creates another complication. Ethernet is strongest as a scale-out technology connecting servers. Proprietary links such as NVLink still play a major role within tightly integrated systems, where bandwidth and latency requirements are even more demanding.

UALink aims to provide an open alternative for scale-up connections. Its progress matters because a buyer can choose Ethernet for scale-out while remaining dependent on a proprietary scale-up fabric. Openness at one network layer does not eliminate concentration elsewhere.

Power consumption is another constraint. Faster switches and more optical links require electricity and cooling. Co-packaged optics promise better efficiency, but they introduce manufacturing, repair, and deployment questions.

Nvidia says its co-packaged optical platform improves network power efficiency compared with pluggable transceivers. Those claims require validation in production systems. Operators will evaluate total power, failure rates, serviceability, and application performance.

The forecast also does not specify which suppliers capture the most value. Switch-system revenue can flow differently from switch-silicon, adapter, cable, and optical revenue. A large market can still produce uneven margins and intense competition.

Finally, the headline error itself is a warning about precision. Market forecasts often move through press releases, publisher summaries, aggregators, and social posts. Each step can remove qualifiers or combine adjacent figures.

Google News is an aggregation layer, not the origin of the underlying market research. Readers should open the publisher’s article and trace important figures to the research firm. A striking number without a defined category and period deserves particular scrutiny.

Three Signals Will Test Dell’Oro’s Forecast

Ethernet deployments, next-generation port adoption, and data center spending will determine whether the forecast remains credible.

The first signal is production adoption of Ultra Ethernet. The specification now exists, but widespread deployment requires shipping hardware, compatible software, and completed interoperability work.

Buyers should watch for named installations that combine equipment from multiple suppliers. Those deployments would support the claim that Ethernet offers practical vendor choice rather than openness only at the specification level.

Performance disclosures will matter more than peak bandwidth. Useful evidence includes job completion time, tail latency, network utilization, failure recovery, and results under congested traffic. Comparisons should use similar clusters and workloads.

A successful wave of interoperable deployments would strengthen Dell’Oro’s Ethernet thesis. Delays, fragmented implementations, or heavy dependence on proprietary extensions would weaken the argument that open Ethernet can displace specialized fabrics broadly.

The second signal is the transition from 800-gigabit ports to 1.6-terabit ports. Dell’Oro expects AI back-end networks to adopt higher speeds earlier than conventional front-end networks. The transition should become visible through switch shipments, optical demand, and customer deployments.

Faster ports can reduce the number of links needed for a given bandwidth, but they also introduce engineering challenges. Signal integrity, optics, thermal density, and cable reach all affect deployment economics.

Nvidia’s Spectrum-X Photonics roadmap offers one test. Its Ethernet platform combines high-speed switching with network adapters and management software. Delivery schedules and named customer installations will show whether the technology moves from announcements into volume operation.

Competing launches from Broadcom-based systems, Arista, Cisco, and other vendors will provide another test. If multiple suppliers ship 1.6-terabit products at scale, price competition and product choice should increase. A narrow supply base would preserve concentration.

Broad adoption of 1.6-terabit ports during 2027 would reinforce Dell’Oro’s spending trajectory. Delays caused by optics, power, qualification, or weak demand would push revenue further into the forecast period.

The third signal is capital spending by hyperscalers and specialized cloud providers. Network demand ultimately follows the construction and expansion of accelerator clusters. Operators must continue ordering servers, power equipment, cooling systems, switches, and optics.

Aggregate capital expenditure is useful, but infrastructure mix matters more. Spending on land, buildings, or power generation does not immediately translate into switch sales. Buyers and investors should separate installed computing capacity from projects that remain under development.

Accelerator utilization offers another clue. High demand and constrained capacity encourage operators to expand clusters. Underused systems or slowing cloud growth can lead them to postpone network purchases.

Earnings reports can also reveal whether networking grows alongside accelerators. Relevant indicators include data center revenue, Ethernet product growth, optical shipments, inventory, and capital commitments from large customers.

Continued data center investment would support the $100 billion switch forecast. Lower spending, delayed facilities, or slower accelerator deployments would weaken it. None of those outcomes would validate the mistaken $1 trillion headline.

The corrected forecast still points to a consequential change. Networks are becoming a larger and more visible part of AI infrastructure, while Ethernet is challenging a specialized fabric that held overwhelming share only a few years ago.

That shift affects more than switch suppliers. It influences cloud competition, accelerator utilization, data center power requirements, and the ability of buyers to combine equipment from different vendors.

The next time a Google News headline attaches a trillion-dollar figure to one infrastructure component, check the category and period first. Then watch real deployments, faster port shipments, and customer spending. Those signals will show whether Dell’Oro’s actual $100 billion outlook is aggressive, conservative, or on target.