Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Tobie Morgan Hitchcock, CEO, SurrealDB
After two years of nonstop AI experimentation, the industry is hitting a wall. Costs are ballooning, architectures are cracking under real-world scale, and early agent pilots are exposing just how hard accuracy and reproducibility really are. In 2026, enterprises will shift from “trying AI” to confronting the hard engineering and economic choices that determine whether these systems can actually deliver value.
These pressures are already reshaping how the most advanced teams think about infrastructure, cloud strategy, and data architecture. Taken together, these constraints are shaping the architectural decisions enterprises must make next. Three developments, in particular, will define that shift.
1. Hyperscale AI goes on-prem
For the last several years, the cloud has been the unquestioned home for AI development. It offered the fastest path to experimentation, elastic scale, and access to GPUs without upfront investment. But as enterprises push into large-scale assistants, copilots, and internal agent platforms, the economics no longer make sense. The biggest adopters are learning what hyperscalers have known all along: GPUs are extraordinarily expensive to run at sustained scale, and cloud pricing models compound those costs quickly.
In 2026, the largest enterprises will begin pulling their most demanding AI workloads back into the datacenter, largely for cost and control decisions. Teams want predictable access to their own GPU estates, and the cloud bills for high-volume AI can approach, or even eclipse, the revenue generated by AI-driven products. At hyperscale, owning or co-locating hardware becomes the more rational long-term strategy.
The market is already adjusting. GPU vendors like NVIDIA are pouring investment in both on-prem enterprise racks and tighter cloud ecosystems. This signals where they believe growth is headed: large enterprises running mission-critical AI on infrastructure they directly control.
This shift won’t apply to everyone; startups and mid-market companies will continue to rely on the cloud for flexibility and speed. But for the top tier, 2026 will mark a decisive turn. Hyperscale AI is moving closer to home, and with it, a new era of architectural ownership is emerging.
2. Cloud costs and architecture complexity will continue to stall adoption
While some enterprises will execute a calculated shift back to owned infrastructure, most organizations will remain stuck in the middle: eager to scale AI, but uncertain about long-term architecture. And whichever direction they look, the price tag is daunting.
Running large AI workloads in the cloud is expensive. Migrating back in-house is even more expensive unless an organization is certain about its future needs, talent availability, and capacity planning. Many teams simply don’t know yet which AI architectures will serve them best in the long term-agent frameworks, vector-first architectures, retrieval pipelines, or multi-model operational stacks.
This uncertainty is becoming one of the biggest inhibitors to AI adoption. Companies know they need to move fast, but they’re stuck trying to determine the “right” technical strategy in a landscape that shifts every quarter. As a result, 2026 will be a year of iterative learning. Teams will continue to lean on cloud providers to experiment quickly, validate hypotheses, and gather data before locking into multi-year commitments.
The appetite to push the boundaries of AI is higher than ever. But until cost models stabilize and architectural patterns consolidate, adoption will be slower and more deliberate than most observers expect.
3. Cracking agent accuracy and reproducibility will become a major challenge – especially in finance
Financial services has been an early leader in Agentic AI experimentation, with dozens of proofs of concept across customer service, compliance workflows, and internal operations. Few have reached mission-critical, autonomous production. The pressure to solve this will intensify in 2026.
The barrier is accuracy and reproducibility. Banks need AI agents that can reason with customers, produce consistent outputs, and provide verifiable answers. But today, responses are often generated by LLMs drawing from multiple databases – DuckDB for analytics, Postgres for transactions, Snowflake for warehousing, InfluxDB for time series, and so on. This creates enormous flexibility, but also introduces opportunities for subtle inconsistencies that undermine trust.
Layer on top the intrinsic limitations of LLMs – their tendency to pattern-match rather than truly understand context – and you get an environment where even small variability becomes unacceptable.
In response to this, next year we’re likely to see an accelerated consolidation of database strategies. Multi-model, AI-native databases will gain traction because they allow teams to represent data from multiple perspectives within a single system. When the underlying data model is unified, context becomes more consistent, and agent reasoning becomes more reliable.
This won’t solve every challenge in agentic AI, but it will give enterprises a pragmatic foundation while we wait for the next leap in model architecture.
2026 will be defined by hard trade-offs: where AI runs, how much it costs, and how accurately it performs in the real world. The enterprises that commit early to clear architectures and reliable data paths will be the ones capable of deploying AI at scale while everyone else is still refining proofs of concept.
##
ABOUT THE AUTHOR
Tobie Morgan Hitchcock is CEO & Co-Founder of SurrealDB. He is an experienced tech entrepreneur, developer, and software engineer, with 17 years’ experience in the software and cloud-computing industries. In 2021 he founded SurrealDB, with the aim of building the ultimate cloud database for tomorrow’s applications. He has experience in a wide range of different software stacks and development languages, with a focus on distributed databases, and highly-available architectures.
