Enterprises are buying agents, not models. Governance is what is holding them back
A Databricks revenue update, a SemiAnalysis spending breakdown, a survey of 150 IT leaders, and a venture firm's infrastructure thesis all point the same way: the buying has started, and the blockers are access, identity, and proof.
- SemiAnalysis allocates 45% of its $15 million annual AI spend to persistent agents, primarily through Perplexity. Self-reported m:…6696
- Databricks reports a $7 billion revenue run rate as of August 2026. Self-reported m:…1745
- 69% of IT and security decision-makers say security concerns are slowing agent adoption; only 27% say their identity systems can govern agents at scale. Self-reported m:…6751
- Permitting, specifically water, wastewater, and air permits, is the primary risk causing gigawatt-scale data center delays in the United States. Self-reported m:…6696
The money moved first
SemiAnalysis now puts 45% of its $15 million annual AI spend on persistent agents, primarily through Perplexity. The same episode reports that Anthropic is offering enterprise discounts to customers who commit to 70% or more of their AI spend with the company, a sign of how hard vendors are competing to lock in the agent budget. m:…6696
Databricks, meanwhile, reports a $7 billion revenue run rate as of August 2026. Its chief executive argues that AGI was achieved last year, based on the consensus definition from the UC Berkeley AMP Lab, and that enterprises are still behind, not because the models fall short but because they lack secure, governed access to internal organizational context. The same episode pushes back on recursive self-improvement, saying it is not currently occurring in AI models. m:…1745
The two accounts agree on where the spending is going even as they disagree about what has been achieved. The budget is flowing to agents that persist and act, not to raw model access, and the vendor fight is over who owns that spend. m:…6696 m:…1745
Databricks also describes using its own medicine internally: an internal agentic platform called Genie that utilizes an ontology graph to automate management tasks. The detail matters because it is the same governed-context argument in miniature. The platform works, on the company's account, because the agents operate over a governed map of the organization's data rather than over raw model capability alone. m:…1745
What is blocking the rest
A January 2026 survey of 150 IT and security decision-makers puts numbers on the blockage: 69% say security concerns are slowing agent adoption, and only 27% agree that their existing identity systems are equipped to govern agents at scale. The survey's authors argue that verifiability is a more important near-term use case for blockchains than payments in the context of AI agents, and that crypto teams can use their experience in signing and delegated authority to help enterprises delegate work to agents while retaining control. m:…6751
The gap between 69% and 27% is the story. Enterprises are not waiting for a smarter model; they are waiting for a way to prove what an agent did, who authorized it, and how to take the authority back. Until identity and audit catch up, the agent budget stays concentrated in the teams willing to run the risk themselves. m:…6751
The infrastructure framing
One venture firm's analysis frames the shift in factory terms. Large enterprises are transitioning to managing AI infrastructure similarly to factories or distribution centers, the authors write, and will adopt a hybrid model combining private capacity and public cloud services. They predict Return on Intelligence will become a primary enterprise metric for measuring AI performance, that software coordinating AI systems will become a strategic control point, and that infrastructure economics will become a dominant factor in enterprise AI strategy over the next five years. m:…0436
The buildout behind that framing has its own bottleneck, and it is not chips. Permitting, specifically water, wastewater, and air permits, is the primary risk causing gigawatt-scale data center delays in the United States, according to SemiAnalysis. And 49% of people in counties with a 100MW+ data center are unaware the facility exists. m:…6696
For investors and operators, the through-line is unglamorous: the winners in this cycle may be decided less by model quality than by governed access, identity, permits, and who controls the coordination layer. Those are procurement and policy problems, and they are where the money is now going. m:…1745 m:…0436
What is still unresolved
- Whether Anthropic's discounts for 70%-plus spend commitments reshape multi-vendor buying or lock it in
- Whether Return on Intelligence becomes a real reporting metric or stays a framing device
- How many enterprises clear the identity and governance bar in the next year
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.