Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Nuno Lopes, Associate Professor at U. Lisbon & Advisor at FuriosaAI
Hot takes about “AI slop,” regulatory battles, and constantly shifting leaderboard rankings will continue to dominate AI headlines. But for enterprises and infrastructure operators, a far more important transition is already underway. The AI industry is moving from a phase of unconstrained experimentation to one governed by economics, physical constraints, and operational reality. By 2026, success in AI will be defined less by who trains the largest model and more by who can deploy AI workloads efficiently, reliably, and profitably. The next wave of innovation will focus on breaking hardware concentration, extracting more value from existing data centers, and using open standards to reduce costs and risk across the stack.
The end of the “No one was ever fired for buying IBM Nvidia” era
The era of single-vendor dependence in AI compute is ending. Supply-chain risk, pricing volatility, and deployment delays are already pushing large buyers toward multi-vendor strategies across accelerators, high-speed memory (HBM and emerging alternatives) and networking. While CUDA remains influential, maturing compiler stacks and standardized model-serving frameworks are steadily lowering switching costs. Moreover, much of CUDA’s moat stems from the availability of engineers fluent in its APIs, a barrier that is rapidly eroding as AI coding assistants make it easier to target unfamiliar hardware and software stacks. For inference workloads in particular, hardware lock-in is no longer inevitable, making diversification not just feasible but economically rational.
AI delivers real ROI in the physical economy
Beyond text and images, AI will increasingly be used to solve NP-hard optimization problems in scheduling, routing, inventory, and pricing. Low-margin industries such as aviation, logistics, manufacturing, and food retail will see disproportionate gains, where even small efficiency improvements translate directly into profit.
Inference becomes the dominant cost center
In 2026, inference will overtake model training as the largest line item for AI compute spending, overtaking training. As models stabilize and reuse increases, competitive advantage will shift toward hardware and infrastructure optimized for sustained, efficient inference rather than peak theoretical throughput.
Efficiency unlocks legacy data centers
As the industry transitions from growth-at-any-cost to sustainable profitability, physical constraints such as power, cooling, and rack density will increasingly dominate AI deployment decisions. In 2026, accelerators that can operate within the limits of existing “brownfield” data centers will see rapid adoption, enabling enterprises to deploy AI immediately rather than waiting years for new facilities or costly retrofits. This is especially true for on-prem data centers in regulated sectors such as healthcare and finance, where high-power-density deployments are often impractical, making efficiency a prerequisite for AI adoption rather than an optimization.
Power efficiency becomes the primary scaling constraint
Power will be the biggest bottleneck for AI infrastructure in 2026, not compute. As inference workloads overtake training in total compute spend, metrics such as performance per watt and queries per joule will matter more than peak throughput. Rising energy costs and limited power availability will force operators to prioritize efficient accelerators that can deliver predictable performance within tight power envelopes, favoring architectures optimized for sustained inference rather than short-lived benchmark performance.
Low-latency inference reshapes deployment economics
Ultra-low latency will become a first-class requirement across more industries, from high-frequency trading and fraud detection to real-time personalization and industrial control systems. Faster inference not only enables new applications but also changes infrastructure economics: cloud providers can place models farther away from end users, in regions with cheaper power or capacity, because faster model response offsets additional network latency. This will favor accelerators designed for consistent, low-latency inference under tight power constraints.
Open standards gain momentum in AI data centers
The steady move toward open standards (Ethernet for networking, PyTorch for training and inference, and frameworks such as vLLM for serving) will accelerate in 2026. These standards reduce operational friction and, more importantly, weaken hardware lock-in by allowing operators to choose accelerators based on efficiency, cost, and deployability rather than proprietary ecosystems. While the industry is still navigating a crowded and sometimes messy software landscape, this period of experimentation has been necessary to identify what works at scale. In 2026, AI infrastructure will begin transitioning toward maturity, with convergence around a smaller set of proven standards, though flexibility and evolution will remain essential.
Load balancers become economic brokers
The next generation of load balancers will actively optimize cost, not just performance. Expect systems that bid for idle or spot capacity across multiple providers, dynamically trading latency headroom for prices up to an order of magnitude lower. These platforms will enforce budget-aware policies, automatically degrading or rejecting workloads when costs exceed predefined thresholds, effectively turning infrastructure orchestration into a financial control layer.
Conclusion
Ultimately, AI’s next phase will look familiar to infrastructure professionals: fewer hype cycles, tighter margins, and a relentless focus on efficiency, interoperability, and ROI. The winners won’t be those with the biggest models, but those with infrastructure that fits into existing data centers, existing budgets, and real-world operational constraints.
##
ABOUT THE AUTHOR
Nuno Lopes has over 20 years of experience building and verifying compilers and toolchains. He is an advisor with FuriosaAI and leads research and teaching at Instituto Superior T�cnico in Lisbon, Portugal. He previously served as a Principal Researcher at Microsoft Research on compiler verification, network verification, and ML-accelerator compilers. His work pairs a PhD’s formal rigor with hands-on engineering across influential open-source projects and industrial ML compiler stacks.
