As 2025 comes to a close, we reflect on several remarkable milestones achieved by the data center switching market this year, and what 2026 may have in store for us.
Looking back at 2025, several clear inflection points reshaped the market:
Ethernet overtakes InfiniBand in AI back-end networking: Supported by strong tailwinds on both the supply and demand sides, 2025 marked a decisive turning point for AI back-end networks, as Ethernet surpassed InfiniBand in market adoption. This shift is particularly striking given that just two years ago, InfiniBand accounted for nearly 80% of the data center switch sales in AI back-end networks.
Overall Ethernet Data Center switch sales nearly doubled compared with 2022: The rapid adoption of Ethernet in AI back-end deployments propelled total Ethernet data center switch sales to an all-time high in 2025, nearly doubling annual revenues compared with 2022 levels.
800 Gbps well surpassed 20 M port within just three years of shipments: As a point of reference, it took 400 Gbps six to seven years to achieve the same milestone
The vendor landscape shifted meaningfully toward AI-exposed players: Vendors with greater exposure to AI back-end networking significantly outperformed the broader market in 2025. Companies such as Accton, Celestica and NVIDIA were among the primary beneficiaries of this shift, reflecting how AI-driven demand is reshaping competitive dynamics. Arista maintained the leading position in the Total Ethernet Data Center Switching market.
Looking ahead to 2026, questions are emerging around whether the pace of investment can be sustained after such an extraordinary year. While skepticism around AI returns on investment is growing, we believe the industry is still in the early innings of a multi-year AI investment cycle. Based on the latest capital expenditure outlooks from the large hyperscalers (Google, Amazon, Microsoft, Meta, Oracle and others), we expect another strong year of AI-related investment in 2026, which should continue to drive robust spending across the networking portion of the infrastructure stack.
Networking is becoming increasingly critical, as it plays a central role in addressing some of the most challenging scaling bottlenecks in AI deployments—including power availability and compute demand. Below are some of the inflection points expected for 2026:
Demand remains exceptionally strong in AI back-end networking. We continue to expect strong double-digit growth in AI networking spending, driven by ongoing scale-out of AI clusters. The integration of co-packaged optics could further accelerate market growth as optics would easily add multi billions to the market size.
Supply constraints remain the primary risk to our forecast. We expect demand to continue to outpace supply, with shortages in chips, memory, and other critical components representing the main caveats to our outlook. As a result, the market remains supply-constrained rather than demand-constrained—a challenging dynamic, but ultimately a more favorable one than the reverse.
Scale-up emerges as a new battlefield for Ethernet. After securing a leading position in the scale-out segment of AI back-end networks, Ethernet is now expanding into scale-up, where NVLink has historically dominated. In this space, Ethernet will compete not only with NVLink but also with UALink, another alternative to NVLink. We anticipate 2026 will be a year full of vendor announcements targeting both Ethernet and UALink opportunities in scale-up. Scale-up represents what could be the largest total addressable market expansion the industry has ever seen.
1.6 Tbps switches expected to ship in volume in 2026. 2026 will mark the first year of volume deployments of 1.6 Tbps switches, driven by the insatiable demand for high bandwidth in AI clusters. 1.6 Tbps ramp is expected to be even faster than 800 Gbps, surpassing 5 M ports within one to two years of shipments.
Co-packaged optics (CPO) expected to ramp on both InfiniBand and Ethernet switches. After many years of development and debate, 2026 is expected to see the initial volume ramp of CPO on both InfiniBand and Ethernet switches. On the demand side, major hyperscalers are actively trialing the technology. On the supply side, while NVIDIA is leading the way, we expect other vendors to follow shortly.
Vendor diversity set to increase in 2026. As AI clusters continue to scale, vendor diversity with both incumbent vendors as well as new entrants, will become increasingly important to ensure risk mitigation and supply availability. We believe that no single vendor can meet the full demand for AI infrastructure. As a result, we expect SONiC adoption to accelerate in both scale-up and scale-out deployments, as it will be critical in enabling this broader vendor ecosystem
In summary, as we look ahead to 2026, the AI-driven data center landscape is set to continue its rapid evolution. From Ethernet’s rise in AI back-end networks and the emergence of scale-up as a new battlefield, to the adoption of 1.6 Tbps switches, co-packaged optics, and a more diverse vendor ecosystem, the infrastructure supporting AI is expanding in both scale and complexity. While supply constraints and ROI questions remain challenges, the industry is clearly in the early innings of a multi-year AI journey. Networking, in particular, will play a pivotal role in enabling the next phase of AI growth, making 2026 an exciting year for both innovation and investment.
[wp_tech_share]
The hyperscale AI infrastructure buildout is entering a more mature phase. After several years of rapid regional expansion driven by resilience, redundancy, and data sovereignty, hyperscalers are now focused on scaling AI compute and supporting infrastructure efficiently. As we move into 2026, the cycle is increasingly defined by capex discipline and execution risk, even as absolute investment levels remain historically high.
Accelerated Servers Remain the Core Spending Driver
Spending on high-end accelerated servers rose sharply in 2025 and continues to anchor AI infrastructure investment heading into 2026. These platforms pull through demand for GPUs and custom accelerators, HBM, high-capacity SSDs, and high-speed NICs and networks used in large AI clusters. While frontier model training remains important, a growing share of deployments is now driven by inference workloads, as hyperscalers scale AI services to millions of users globally.
This shift meaningfully expands infrastructure requirements, as inference workloads require higher availability, geographic distribution, and tighter latency guarantees than centralized training clusters.
GPUs Continue to Dominate Component Revenue
High-end GPUs will remain the largest contributor to component market revenue growth in 2026, even as hyperscalers deploy more custom accelerators to optimize cost, power efficiency, and workload-specific performance at scale. NVIDIA is expected to begin shipping the Vera Rubin platform in 2H26, which increases system complexity through higher compute and networking density and optional Rubin CPX inference GPU configurations, materially boosting component attach rates.
AMD is positioning to gain share with its MI400 rack-scale platform, supported by recently announced wins at OpenAI and Oracle. Despite growing competition, GPUs continue to command outsized revenue due to higher ASPs, broader ecosystem support.
Near-Edge Infrastructure Becomes Critical for Inference
As AI inference demand accelerates, hyperscalers will need to increase investment in near-edge data centers to meet latency, reliability, and regulatory requirements. These facilities—located closer to population centers than centralized hyperscale regions—are essential for real-time, user-facing AI services such as copilots, search, recommendation engines, and enterprise applications.
Near-edge deployments typically favor smaller but highly dense accelerated clusters, with strong requirements for high-speed networking, local storage, and redundancy. While these sites do not approach the power scale of centralized AI campuses, their sheer number and geographic dispersion represent a meaningful incremental capex requirement heading into 2026. In contrast, far-edge deployments remain more use-case dependent and are unlikely to see material growth until ecosystems and application demand further mature.
Networking and CPUs Transition Unevenly
The x86 CPU and NIC markets tied to general-purpose servers are expected to decelerate in 2026 following short-term inventory digestion. In contrast, demand for high-speed networking remains tightly linked to accelerated compute growth. Even as inference workloads outpace training, inference accelerators continue to rely on scale-out fabrics to support utilization, redundancy, and ultra-low latency.
Supply Chains Tighten as Component Costs Rise
AI infrastructure supply chains are becoming increasingly constrained heading into 2026. Memory vendors are prioritizing production of higher-margin HBM, limiting capacity for conventional DRAM and NAND used in AI servers. As a result, memory and storage prices are rising sharply, increasing system-level costs for accelerated platforms.
Beyond memory, longer lead times for advanced substrates, optics, and high-speed networking components are adding further volatility to the supply chain. In parallel, tariff uncertainty and evolving trade policy introduce additional supply-chain risk, and potentially elevating component pricing over the medium term.
Capex Remains Elevated, but ROI Scrutiny Intensifies
The US hyperscale cloud service providers continue to raise capex guidance, reinforcing the continuity of the multi-year AI investment cycle into 2026. Accelerated computing, greenfield data center builds, near-edge expansion, and competitive pressures remain strong tailwinds. Changes in depreciation treatment provide levers to optimize cash flow and support near-term investment levels.
However, infrastructure investment has outpaced revenue growth, increasing scrutiny around capex intensity, depreciation, and long-term returns. While cash flow timing can be managed, underlying ROI depends on successful AI monetization, increasing the risk of margin pressure if revenue growth lags infrastructure deployment.
[wp_tech_share]
A year of continuous shifts within the sector — and familiar debates beyond it
As we close out 2025, the milestones of the past twelve months underscore just how quickly the industry is shifting beneath our feet. DeepSeek’s breakthrough reshaped assumptions about compute efficiency and cost; NVIDIA’s announcement of Blackwell Ultra signaled yet another leap in accelerator performance; the White House’s AI Action Plan formalized the policy stakes around national compute capacity; Stargate’s Abilene facility began operating at unprecedented scale, becoming a symbol of the AI‑era mega‑campus; debates around AI circular investments highlighted both the ambition and fragility of capital flows into frontier infrastructure — only to name a few of key milestones of this past year
These developments set the stage for a year that will balance continuity with disruption. For vendors and operators, 2026 will bring meaningful shifts in technologies, architectures, and competitive dynamics. Yet from the outside, the narrative may feel familiar. The same themes that began surfacing more prominently in recent years — and defined public debate throughout 2025 — will continue to dominate headlines, even as the underlying infrastructure evolves at a far faster pace.
What we’re not predicting — because everyone else already is
Power scarcity remains the defining constraint, with power availability continuing to be the single most important determinant of site selection for data center projects. Speculation about an AI‑driven investment bubble is expected to intensify, as trillions of dollars in critical infrastructure are deployed amid lingering uncertainty about long‑term monetization models. And public visibility of the sector will keep rising, bringing sharper community pushback, permitting resistance, and societal concerns ranging from energy affordability to the impact of AI on jobs, as well as growing scrutiny over the safe and responsible use of AI, particularly among young people — pressures that intensify most as the industry lacks coherent, accessible, and positive messaging about its value to communities and the broader economy.
Because these forces are so obvious and so deeply embedded in the industry’s trajectory, we will not include them among our predictions. Instead, this outlook focuses on the emerging dynamics that will shape vendors, operators, and the broader ecosystem in ways both expected and unexpected.
The easy ones: our highest-confidence expectations for 2026
These trends are already well underway, with early signals evident throughout 2025, reinforcing a trajectory that leaves little doubt about their momentum heading into 2026.
1. Consolidation and partnerships accelerate
The complexity of gigawatt‑scale data centers is pushing vendors to work together more closely, driving a surge in strategic partnerships that combine expertise across power, cooling, controls, and integration. Expect more joint reference architectures, co‑engineered solutions, and collaborative designs that extend well beyond any single vendor’s historical domain. We anticipate at least ten additional partnership announcements in 2026 as vendors align to meet the growing demands of AI‑era infrastructure.
In parallel, consolidation will continue as vendors with differentiated capabilities become acquisition targets — particularly in high-priority areas such as liquid cooling, solid-state power electronics, and global design and service expertise. These acquisitions will further accelerate the shift toward full-stack delivery models, with integrated chip-to-rack, rack-to-row, and row-to-hall solutions becoming a defining competitive strategy. We expect no fewer than five acquisitions or take-private transactions crossing the $1 billion threshold, underscoring the intensifying race to secure critical capabilities across the DCPI stack.
2. Real builds matter more than bold visions (and vanished ones)
Multi‑billion‑dollar and multi‑gigawatt campus announcements might continue to dominate headlines, but the center of gravity will shift toward execution rather than ideation. Operators will focus on translating these bold visions into reality — securing power, navigating permitting, sequencing construction, and commissioning facilities on time.
With the running backlog of public announcements now exceeding 70 GW of stated capacity, a meaningful share of these projects is likely to remain “braggerwatts” — aspirational declarations that never progress past land options, concept designs, or early‑stage filings. As economic, regulatory, and power‑availability constraints sharpen, attention will shift back to credible projects with clear pathways to completion and well‑defined delivery plans.
Today, several sites are on trajectories that suggest they could eventually cross the fabled 1 GW capacity threshold, but none have reached that milestone yet. By the end of 2026, however, we expect at least five sites worldwide to surpass 1 GW of operational capacity.
3. Divergence grows before convergence returns
Despite efforts toward convergence, 2026 is likely to bring even greater architectural divergence across power and cooling, a proliferation of design pathways rather than a narrowing of them. This is being fueled by rapid technological shifts that show no signs of slowing.
On the power side, even as clarity improves around 400 Vdc and 800 Vdc rack architectures, vendors will diversify rather than narrow their portfolios — developing new families of DC circuit breakers, power shelves, hybrid and supercapacitor‑based energy storage, and MV switchgear integrated with solid-state electronics in preparation for deployments expected in 2028/29.
Cooling will see similar diversification. A testing ground of novel technologies — including two‑phase direct liquid cooling (DLC), CDU‑less single‑phase DLC, and a wide variety of cold‑plate architectures — is expected to gain momentum, expanding the solution diversity of the ecosystem.
In this environment, initiatives like the Open Compute Project (and its collaborations with ASHRAE, Current/OS, and others) will become even more important in steering the industry, offering reference frameworks and shared direction to help channel innovation while reducing unnecessary fragmentation.
Watch closely: trends gaining momentum — but not yet locked in
Early signals suggest these trends could gain real traction — but timing, economics, and scale remain uncertain.
4. “Micro‑mega” edge AI deployments are on the rise
As compute density within a single rack skyrockets, many AI workloads will be able to operate on one — or just a handful — of cabinets. These compact yet powerful clusters will increasingly sit alongside conventional compute to support hybrid workloads. Expect a wave of megawatt-class, ultra-dense AI racks for enterprise post-training and inference — small-scale AI factories — embedded within colocation sites, enterprise campuses, or telco edge facilities.
What makes this shift noteworthy is what it reveals about broader AI adoption: AI is moving beyond pilots and proofs‑of‑concept and into day‑to‑day business operations, requiring right‑sized, high‑density compute footprints placed directly where data and decision‑making occur.
Architecturally, this marks a meaningful shift. Instead of concentrating accelerated compute solely in hyperscale campuses or purpose‑built training clusters, enterprises and colocators will increasingly deploy AI directly into existing facilities. This proximity to business‑critical workflows will drive demand for modular, pre‑engineered AI systems that can be “dropped in” with minimal disruption, along with managed AI‑infrastructure services that oversee monitoring, lifecycle management, and performance optimization.
5. Air cooling strikes back
The novelty of liquid cooling has dominated industry discourse for the past three years, pushing vendors and operators to rapidly adapt — bringing new products to market, redesigning systems to accommodate liquid infrastructure, and upskilling operational teams to support deployments at scale. But as AI deployments move beyond frontier‑model training clusters and into enterprise environments, high‑density AI racks will more frequently appear in facilities not originally designed for liquid cooling.
This shift will prompt a resurgence in advanced air‑cooling solutions. Expect a proliferation of 40–80 kW air‑cooled racks supported by extremely high‑performance thermal systems, paired with 60–150 kW liquid‑cooled racks equipped with liquid‑to‑air sidecars. The result: hybrid thermal profiles within the same facility, introducing complex challenges for operators managing uneven heat loads and airflow dynamics.
Far from being overshadowed by liquid cooling, air‑cooling systems are poised for incremental growth as operators seek flexible, retrofit‑friendly approaches to support heterogeneous rack densities across mixed‑use sites.
6. Immersion cooling re-emerges in modular form
After the hype cycle of recent years, immersion cooling is beginning to find its footing in more targeted, pragmatic applications. Rather than competing head‑on with DLC for hyperscale AI clusters, immersion vendors are shifting toward modular, compact systems that deliver differentiated value.
We expect growing traction in edge, telecom, and industrial environments, where immersion’s sealed‑bath architecture offers advantages in reliability, environmental isolation, and minimal site modification. These deployments will remain modest in scale, but meaningful in carving out a sustainable niche beyond today’s supercomputing and crypto segments.
To be clear, immersion cooling is not poised to displace DLC or become a dominant cooling technology. However, it is finally entering a phase where use‑cases align with its strengths — enabling vendors to build viable businesses around modular, ready‑to‑deploy immersion clusters that “drop in” alongside traditional IT and support workloads that benefit from simplified thermal management and rapid deployment.
7. Europe and China wake up — but in very different ways
Europe and China are both poised for stronger AI‑driven data‑center momentum in 2026, but their trajectories could not be more different. In power‑constrained Europe, growth will increasingly hinge on inference deployments located closer to population centers, to minimize network latency (even if compute latency remains the bigger challenge for AI services). This shift toward user‑proximate infrastructure will steer investment toward distributed, high‑density nodes rather than massive gigawatt-scale training campuses. Within this landscape of smaller facilities, a growing cohort of start‑up model builders will prioritize hyper‑efficient architectures that can extract maximum utility from these distributed fleets, for both inference and selective training workflows.
China, by contrast, faces no shortage of power. Its constraint is access to the latest generation of advanced accelerators. We expect operators to continue building at scale using a mix of domestic silicon and whatever Western supply remains available — iterating rapidly as local manufacturers improve capability generation by generation. Over the next few years, this mix‑and‑match strategy will help China bridge the gap until it achieves greater semiconductor self‑sufficiency, resulting in substantial expansion of AI data‑center capacity even under export controls.
The long shots: unlikely swings with outsized impact
Three low-probability but transformative developments, if they emerge, could reshape the data center landscape far more than their probability suggests.
8. U.S. government tightens regulation of the data center industry
A push in Washington to encourage investment in advanced cooling technologies — including a proposed bill aimed at accelerating liquid‑cooling adoption — could have unintended consequences. While well‑intentioned, efforts to steer technological choices risk drawing the federal government more directly into data center design decisions, increasing oversight and potentially making infrastructure requirements more rigid at a time when flexibility is essential.
We do not expect sweeping regulation to materialize in 2026. The current administration has closely aligned itself with AI as a pillar of economic competitiveness and will be wary of stymieing data center buildout, especially given its role in supporting GDP growth. Moreover, political attention will be dominated largely by the mid‑term elections, leaving little bandwidth for complex industry‑specific legislation.
However, affordability and household cost pressures are set to become highly charged political themes — and in that environment, data centers may attract negative scrutiny. As utilities grapple with rising demand and public concern around bills, the industry could face a wave of unfavorable headlines and heightened calls for transparency. To mitigate reputational risk, operators will need to invest more heavily in public engagement, clear messaging, and proactive demonstration of their contributions to reliability, economic growth, and community well‑being.
9. The first liquid-cooling leak critical failure hits the headlines
The early wave of liquid-cooled deployments often moved faster than the industry’s collective design and operational expertise. Many systems were installed without fully accounting for the nuances of coolant management, materials compatibility, monitoring, and routine maintenance — conditions that naturally elevate leak risk. Throughout 2025, we saw scattered reports of cluster-level shutdowns tied to liquid-handling failures, but nothing approaching the scale or societal visibility of a major cloud outage.
While we still believe high-profile failures are possible, their broader impact will likely be limited. Despite growing enterprise adoption, most AI systems are not yet embedded deeply enough in critical business processes to trigger widespread disruption. As a result, even a significant leak-related outage is unlikely to spark the kind of global headlines seen after the AWS blackout — though it may accelerate industry efforts around standards, training, instrumentation, and risk-mitigation practices.
10. The GPU secondary market skyrockets
As hyperscalers and neo cloud providers refresh their fleets, early generations of GPUs — notably Ampere- and Hopper-based accelerators — will increasingly face retirement to make room for newer, more efficient architectures. This raises a key question already weighing on investors: what is the real depreciation timeline for AI hardware on hyperscaler balance sheets?
We expect most older GPUs to shift into lower‑complexity inference workloads or the training of smaller, less compute‑intensive models. We believe it is still too early for widespread scrapping of full data centers built on these platforms, which could flood the secondary market of GPUs looking for another productive life somewhere else.
Enterprise IT environments and colocation providers will see growing volumes of these second‑hand GPUs entering their ecosystems, often at attractive price points. Integrating these “intruders” into general‑purpose, lower‑density compute environments will introduce new operational and thermal challenges. Operators will need to manage concentrated heat loads, non‑uniform rack densities, and power profiles that differ from their conventional estate.
The bubbling question we can’t avoid — even if we tried
Speculation about an AI “bubble” has increasingly dominated media narratives throughout 2025, and the conversation is unlikely to quiet down in 2026. It is true that many AI‑adjacent companies are trading at lofty valuations, buoyed by optimism around future adoption and monetization, an optimism may not prove durable. There is a meaningful possibility that equity markets enter correction territory in 2026, bringing P/E ratios closer to historical norms.
Yet even in a cooling market environment, we do not expect the data‑center buildout to slow materially. Hyperscalers continue to generate ample cash flow to support aggressive infrastructure expansion, and their balance sheets remain low‑leveraged, giving them capacity to secure additional capital if needed. Strategic imperatives will outweigh short‑term market pressure: these companies are locked in a race to establish AI hegemony — or risk being left behind.
In other words, financial markets may wobble, but the underlying drivers of AI infrastructure investment remain intact. The bubble debate will rage on, but the buildout will continue.
Looking ahead: embracing another year of acceleration and uncertainty
As with every prediction cycle, only time will reveal which of these dynamics take hold and which fade into the background. What is certain, however, is that 2026 will yet again challenge our assumptions. The pace of AI‑driven infrastructure evolution shows no signs of slowing, and the industry will continue navigating a rare combination of technological disruption, supply‑chain reinvention, and unprecedented demand for capacity.
While we avoid grand year‑end platitudes, it is fair to say that much will change — and much will stay the same. Power will remain the currency of competitiveness, AI will continue to push infrastructure to its limits, and operators and vendors alike will be forced to adapt faster than ever. At Dell’Oro Group, we look forward to tracking, analyzing, and interpreting these shifts as they unfold.
Here’s to a 2026 that will undoubtedly keep all of us in the data‑center world busy — and to the insights that the next twelve months will bring!
[wp_tech_share]
AI capacity announcements are multiplying fast—but many overlap, repeat, or overstate what will realistically be built. Untangling this spaghetti means understanding when multiple headlines point to the same capacity and recognizing that delivery timelines matter as much as the billions of dollars and gigawatts announced.
AI is often hailed as a force set to redefine productivity — yet, for now, it consumes much of our time simply trying to make sense of the scale and direction of AI investment activity. Every week brings record-breaking announcements: a new model surpassing benchmarks, another multi-gigawatt data center breaking ground, or one AI firm taking a stake in another. Each adds fuel to the frenzy, amplifying the exuberance that continues to ripple through equity markets.
When AI Announcements Become “Spaghetti”
In recent weeks, the industry’s attention has zeroed in on the tangled web of AI cross‑investments, often visualized through “spaghetti charts.” NVIDIA has invested in its customer OpenAI, which, in turn, has taken a stake in AMD — a direct NVIDIA competitor — while also becoming one of AMD’s largest GPU customers. CoreWeave carries a significant investment from NVIDIA, while ranking among its top GPU buyers, and even leasing those same GPUs back to NVIDIA as one of its key compute suppliers. These overlapping stakes have raised questions about governance and prompted déjà vu comparisons with past bubbles. Morgan Stanley’s Todd Castagno captured this dynamic in his now‑famous spaghetti chart, featured in Barron’s and below, which quickly circulated among investors and analysts alike.
Source: Morgan Stanley
Why Venn Diagrams Matter More Than Spaghetti Charts
Yet while investors may have reason to worry about these tangled relationships, data center operators, vendors, and analysts should be paying attention to two other kinds of charts: Venn diagrams and Gantt charts.
In our conversations at Dell’Oro Group’s data center practice, we’re consistently asked: “How much of these announced gigawatts are double‑counted?” and “Can the industry realistically deliver all these GWs?” These are the right questions. For suppliers trying to plan capacity and for investors attempting to size the real opportunity, understanding overlap is far more important than tracking every new headline.
When all public announcements are tallied, the theoretical pipeline can easily stretch into the several‑hundred‑gigawatt range — far above what our models suggest will actually be built by 2029. This leads to the core issue: how do we make sense of all these overlapping (and at times even contradicting) announcements?
The OpenAI Example: One Company, Multiple Overlapping GW Claims
Its flagship campus in Abilene, Tex. is part of Crusoe’s and Lancium’s Clean Campus development, expected to provide about 1.2 GW of that capacity. The initiative also includes multiple Oracle‑operated sites totaling around 5 GW (including the Crusoe-developed Abilene project, which Oracle will operate for OpenAI, and other sites developed with partners like Vantage Data Centers), plus at least 2 GW in leased capacity from neocloud provider CoreWeave. That leaves roughly 3 GW of U.S. capacity yet to be allocated to specific data center sites.
Assuming Stargate’s full 10 GW materializes domestically, OpenAI’s remaining 16 GW from its 26 GW of chip‑related announcements is still unallocated to specific data center projects. A portion of this may be absorbed by overseas Stargate offshoots in the U.A.E., Norway, and the U.K., generally developed with partners such as G42 and Nscale. These countries are already confirmed locations, but several additional European and Asian markets are widely rumored to be next in line for expansion.
While OpenAI‑dedicated Stargate sites draw significant attention, the reality is that most of the remaining capacity likely ties back to Microsoft — the model builder’s largest compute partner and major shareholder. Microsoft’s new AI factories, including the Fairwater campus in Wisconsin, have been publicly described as shared infrastructure supporting both Microsoft’s own AI models and OpenAI’s workloads.
Naturally, Microsoft’s multibillion‑dollar capex program has come under close investor scrutiny. But to understand actual capacity expansion, one must ask: how much of this spend ultimately supports OpenAI? Whether through direct capital commitments or via absorbed costs within Azure‑hosted AI services, a meaningful share of Microsoft’s infrastructure buildout will inevitably carry OpenAI’s workloads forward.
Given the size and complexity of these projects, it’s unsurprising that multiple stakeholders — chipmakers, cloud providers, developers, utilities, and investors — announce capacity expansions tied to the same underlying sites.
A clear example is Stargate UAE, which has been unveiled from multiple angles:
Each announcement, viewed in isolation, can sound like a separate multi‑gigawatt initiative. In reality, they describe different facets of the same underlying build. And importantly, this is not unique to Stargate — similar multi‑angle, multi‑announcement patterns are becoming increasingly common across major AI infrastructure projects worldwide. This layered messaging contributes to a landscape where genuine incremental expansion becomes increasingly difficult to differentiate from multiple narratives referring to the same capacity.
Source: Dell’Oro Group’s Analysis
Beware the Rise of “Braggerwatts”
If tracking real, shovel‑ready projects weren’t already challenging enough, a newer phenomenon has emerged to further distort expectations: “braggerwatts.”
These headline‑grabbing gigawatt declarations tend to be bold, aspirational, and often untethered from today’s practical constraints. They signal ambition more than bankability. While some may eventually break ground, many originate from firms without sufficient financing — or without the secured power required to energize campuses of this scale. In fact, the absence of power agreements is often the very reason these announcements become braggerwatts: compelling on paper, but unlikely to materialize.
Power is the Real Constraint—Not Chips
This leads directly to the most consequential source of uncertainty: power. As Microsoft CEO Satya Nadella put it in BG2 podcast, “You may actually have a bunch of chips sitting in inventory that I can’t plug in … it’s not a supply issue of chips; it’s actually the fact that I don’t have warm shells to plug into.”
Recent reports from Santa Clara County, Calif. underscored this reality. Silicon Valley Power’s inability to energize new facilities from Digital Realty and STACK Infrastructure revealed just how fragile power‑delivery timelines have become. Developers, competing for scarce grid capacity, increasingly reserve more power across multiple markets than they ultimately intend to use. Nicknamed “phantom data centers” by the Financial Times, these speculative reservations may be a rational hedging strategy — but they also clog interconnection queues and introduce yet another form of double counting.
Gantt Charts and Reality Checks
Making sense of real data center capacity — especially when announced timelines often compress multi‑year build cycles into optimistic one‑ or two‑year horizons — is challenging enough, but an even bigger issue is that, while announcements are rich in dollars and gigawatts, they are often strikingly vague as to when this capacity will actually be delivered. Several large AI‑era projects have publicized increasingly compressed “time‑to‑token” goals.
Recent mapping by nonprofit Epoch.AI, below, illustrates highly ambitious timelines to the first gigawatt of capacity. Yet the reality is far more measured. Most hyperscale and AI‑focused campuses are expected to phase in capacity over multiple years to manage engineering complexity, navigate permitting, and align with the risk tolerance of investors financing these developments.
Source: EPOCH AI
True Modelling Requires Ground-true Data—Not Hype
Ultimately, this creates a disconnect between what is announced and what is genuinely achievable. Understanding true data center growth requires cutting through overlapping announcements, aspirational gigawatt claims, and speculative power reservations. By grounding expectations in semiconductor shipment volumes, verifiable construction progress, and secured power commitments, the industry can move beyond headline noise and toward an accurate view of the capacity that is truly on the way.
[wp_tech_share]
Across hyperscalers and sovereign clouds alike, the race is shifting from just model supremacy to infrastructure supremacy. The real differentiation is now in how efficiently GPUs can be interconnected and utilized. As AI clusters scale beyond anything traditional data center networking was built for, the question is no longer how fast can you train? but can your network keep up? This is where emerging architectures like Optical Circuit Switches (OCS) and Optical Cross-Connects (OXC), a technology used in wide area networks for decades, enter the conversation.
The Network is the Computer for AI Clusters
The new age of AI reasoning is ushering in three new scaling laws—spanning pre-training, post-training, and test-time scaling—that together are driving an unprecedented surge in compute requirements. At GTC 2025, Jensen Huang stated that demand for compute is now 100× higher than what was predicted just a year ago. As a result, the size of AI clusters is exploding, even as the industry aggressively pursues efficiency breakthroughs—what many now refer to as the “DeepSeek moment” of AI deployment optimization.
As the chart illustrates, AI clusters are rapidly scaling from hundreds of thousands of GPUs to millions of GPUs. Over the next five years, the expectation is that there will be about 124 gigawatts of capacity to be brought online, or an equivalent of more than 70 million GPUs to be deployed. In this reality, the network will play a key role in connecting those GPUs in the most optimized, efficient way. The network is the computer for AI clusters.
Challenges in Operating Large Scale AI Clusters
As shown in the chart above, the number of interconnects scales exponentially with the number of GPUs. This rapid increase drives significant cost, power consumption, and latency. It is not just the number of interconnects that is exploding—the speed requirements are rising just as aggressively. AI clusters are fundamentally network-bound, which means the network must operate at nearly 100 percent efficiency to fully utilize the extremely expensive GPU resources.
Another major factor is the refresh cadence. AI back-end networks are refreshed roughly every two years or less, compared to about five years in traditional front-end enterprise environments. As a result, speed transitions in AI data centers are happening at almost twice the pace of non-accelerated infrastructure.
Looking at switch port shipments in AI clusters, we expect the majority of ports in 2025 will be 800 Gbps. By 2027, the majority will have transitioned to 1.6 Tbps, and by 2030, most ports are expected to operate at 3.2 Tbps. This progression implies that the data center network’s electrical layer will need to be replaced at each new bandwidth generation—a far more aggressive upgrade cycle than what the industry has historically seen in front-end, non-accelerated infrastructure.
The Potential Role of OCS in AI Clusters
Optical Circuit Switches (OCS) or Optical Cross-Connects (OXC) are network devices that establish direct, light-based optical paths between endpoints, bypassing the traditional packet-switched routing pipeline to deliver near-zero-latency connectivity with massive bandwidth efficiency. Google was the first major hyperscaler to deploy OCS at scale nearly a decade ago, using it to dynamically rewire its data center topology in response to shifting workload patterns and to reduce reliance on power-hungry electrical Ethernet fabrics.
A major advantage of OCS is that it is fundamentally speed-agnostic—because it operates entirely in the optical domain, it does not need to be upgraded each time the industry transitions from 400 Gbps to 800 Gbps to 1.6 Tbps or beyond. This stands in stark contrast to traditional electrical switching layers, which require constant refreshes as link speeds accelerate. OCS also eliminates the need for optical-electrical-optical (O-E-O) conversion, enabling pure optical forwarding, that not only reduces latency but also dramatically lowers power consumption by avoiding the energy cost of repeatedly converting photons to electrons and back again.
The combined benefit is a scalable, future-proof, ultra-efficient interconnect fabric that is uniquely suited for AI and high-performance computing (HPC) back-end networks, where east-west traffic is unpredictable and bandwidth demand grows faster than Moore’s Law. As AI workload intensity surges, OCS is being explored as a way to optimize the network.
OCS is a Proven Technology
Using an OCS in a network is not new. It was, however, called by different names over the past three decades: OOO Switch, all-optical switch, optical switch, and optical cross-connect (OXC). Currently, the most popular term for these systems used in data centers is OCS.
It has been used in the wide area network (WAN) for many years to solve a similar problem set. And for many of the same reasons, tier-one operators worldwide have addressed it through the strategic use of OCSs. Hence, OCSs have been used in carrier networks by operators with the strictest performance and reliability requirements for over a decade. Additionally, the base optical technologies, both MEMS and LCOS, have been widely deployed in carrier networks and have operated without fault for even longer. Stated another way, OCS is based on field-proven technology.
Whether used in a data center or to scale across data centers, an OCS offers several benefits that translate into lower costs over time.
To address the specific needs for AI data centers, companies have launched new OCS products. The following is a list of the products available in the market:
Final Thought
AI infrastructure is diverging from conventional data center design at an unprecedented pace, and the networks connecting GPUs must evolve even faster than the GPUs themselves. OCS is not an exotic research architecture; it is a proven technology that is ready to be explored and considered for use in AI networks as a way to differentiate and evolve them to meet the stringent requirements of large AI clusters.