This is our second quarterly update on the state of the AI infrastructure industry. In our first competitive brief, published in September 2025, we segmented the industry into groups of like-minded companies and created our own lifecycle-phase diagram. We argued that many market reports to date have failed to capture the true competitive dynamics, for example, by lumping companies like DigitalOcean, Baseten, and CoreWeave into the same bucket despite their very different business models and value propositions.
Our goal remains to publish a competitive brief each quarter to summarize recent disruptions, trends, and emerging startups and to track how the industry evolves. In this snapshot, we have updated our diagram to reflect new public developments and rearranged some segments where warranted.
Current State: Macro Backdrop & AI Infrastructure Resilience
The broader economy is under stress. Many sectors are contracting, layoffs continue, and consumer demand is muted.
At the same time, a subset of companies tied to AI infrastructure, cloud platforms, and enterprise AI deployments continues to grow. Competition between specialized hardware providers and general-purpose infrastructures is accelerating as demand remains strong.
Despite macroeconomic headwinds, the need for AI compute, both for training and inference, remains very high. This bifurcation has created a divergence where some technology verticals stagnate while AI-related sectors expand.
We are in a recession, but not all industries are affected equally. AI infrastructure is holding up, and in some cases accelerating, despite broader economic softness.
This raises the question: Is the AI bubble about to pop?
Our view is that a sudden collapse is unlikely. A pop would imply a sharp reversal. Instead, we anticipate a gradual recalibration over the next one to two years as consolidation, shifting business models, and new competitive patterns play out. Rising demand for AI hardware, software, and infrastructure across enterprises, governments, and cloud providers supports this outlook.
Key Developments
Google vs. Nvidia: The AI-Hardware Turf War Heats Up
Google has pushed its Tensor Processing Unit (TPU) business into the commercial spotlight. Although TPUs have been used internally for years, recent reports suggest that Google intends to sell or rent TPU capacity more broadly, placing it in direct competition with Nvidia’s GPU ecosystem.
TPUs are increasingly pitched as a lower-cost, inference-optimized alternative to GPUs. Some operators claim that migrating large inference workloads from GPUs to TPUs can yield four to five times cost efficiency and significant power savings.
TPUs, however, are not a full GPU replacement. Training workloads and the need for broad software support still favor GPUs. Even so, the AI-hardware market is shifting from a GPU-centric model to a multi-platform future that includes GPUs, ASICs, TPUs, and custom accelerators.
Amazon Web Services (AWS) Launches AI Factories, On-Prem AI Infrastructure Arrives
AWS recently introduced AWS AI Factory, a solution that allows enterprises and government organizations to deploy dedicated AI infrastructure inside their own data centers. This includes GPUs, Trainium accelerators, networking, storage, databases, and AI services.
These deployments function like private AWS regions, giving customers control over data locality, compliance, and security while retaining tight integration with the AWS AI stack.
This shift reflects a growing reality. Not all organizations can rely solely on public cloud AI due to data-sovereignty requirements, latency constraints, or regulatory pressures. Hybrid and on-prem solutions are becoming mainstream.
Cloudflare Acquires Replicate, Democratising AI Deployment
Cloudflare announced its acquisition of Replicate in November 2025. Replicate offers a large marketplace of open-source and commercial AI models that can be deployed with minimal configuration.
Cloudflare plans to integrate this marketplace, which includes more than fifty thousand models, into its global edge network. This will allow developers to run complex inference workloads closer to users with minimal boilerplate.
The acquisition highlights a broader trend toward abstraction. As hardware becomes more specialized and complex, companies are seeking to hide that complexity behind clean APIs and developer-friendly tooling.
Nvidia’s VRAM-Bundling Drop: GPU Supply Chain Shaken Amid Memory Crunch
According to reports, Nvidia has stopped bundling VRAM with GPU dies sold to add-in-board partners. A global memory supply crunch, driven by AI data center demand, appears to be the cause.
This shift could lead to fewer consumer GPU launches, higher prices, and tighter supply as manufacturers scramble to source memory independently. It also signals a reconfiguration of the GPU supply chain that may accelerate consolidation or shift more demand toward AI-optimized hardware.
Competitive Landscape
The competitive landscape is in constant upheaval, and major disruptions now occur on a quarterly basis. This volatility is most visible in the startup ecosystem. A steady flow of new companies is emerging from stealth with sizable funding. These companies push innovation forward while large incumbents such as CoreWeave slow down as they scale.
The first wave of companies focused on GPU as a Service, offering raw compute capacity. The next wave looks very different. A growing set of software-first startups is building compilers, runtimes, scheduling layers, caching systems, memory optimizers, inference accelerators, observability suites, and orchestration tools. These components are becoming essential as model sizes, context lengths, and hardware mixes evolve unpredictably.
We expect the AI software stack to evolve for at least two more years. Dozens of components still need to mature before the ecosystem reaches any type of equilibrium. During this period, the landscape will continue to expand horizontally with new categories and vertically as vendors move up and down the stack to differentiate.
Outlook & Risks for Next 1–2 Quarters
Expect continued fragmentation across the hardware layer as GPUs, TPUs, ASICs, and hybrid solutions co-exist. Optimization will vary by workload type, cost structures, energy constraints, latency targets, and data-locality requirements.
Enterprise and public sector demand for on-prem and sovereign AI solutions will rise as regulatory and compliance pressures increase. This may accelerate the growth of AI-factory providers, managed hybrid-infrastructure vendors, and system integrators.
The abstraction layer will gain traction. Model deployment platforms, edge AI frameworks, and serverless AI runtimes will make AI more accessible to developers. Competitive advantage may shift toward companies that control this layer rather than those that only provide hardware.
Risks remain. Memory and component supply constraints, power demands for large-scale clusters, economic slowdown affecting enterprise AI spending, and uncertainty about which hardware and software stacks will dominate all create real pressure.
Conclusion
AI infrastructure continues to evolve at a pace unmatched by most other technology sectors. The market is no longer defined by a single axis such as GPU capacity or cloud availability. Instead, it is shaped by competing hardware philosophies, new software layers, global supply pressures, regulatory demands, and a startup ecosystem that produces new categories every quarter.
The coming year will bring more volatility, more innovation, and more consolidation. The companies that succeed will be those that can adapt to shifting hardware topologies, rising energy constraints, and an increasingly complex software stack. For now, the sector shows no signs of slowing. The next snapshot will likely document another series of disruptions and expansions as AI infrastructure continues its rapid transformation.