.webp)
We just got back from San Francisco, a week split between HumanX 2026 and the Apache Iceberg Summit at the Marriott Marquis.
At HumanX, AWS CEO Matt Garman opened the week with a sharp observation: the real barrier to agentic AI isn't the technology. It's accountability, how organizations structure decision-making when the decision-maker is a machine. The floor agreed. Nobody was debating whether to implement AI anymore. The conversation moved to how you scale it without losing control.
One moment from the week stuck with me more than any keynote. I got invited to a private dinner, the kind where the invite list actually matters. What I didn't know until I arrived: the guest list was curated by an LLM. An agent had gone through attendee profiles and decided who was worth inviting. I made the cut, which was exciting!
The dinner was genuinely great. But at some point someone asked the question out loud that none of us could answer: how did it decide? What data did it look at? What was it optimizing for? Could anyone pull up a log and actually check?
Silence.
That question followed us through every conversation for the rest of the week, at HumanX sessions on agent governance, on the expo floor, and across town at the Iceberg Summit. And every time we asked, the honest answer was the same: no, we can't audit it.
Agent Observability Was the Buzzword. But It Only Goes Halfway.
The hottest topic at HumanX was agent observability. AI SRE stacks were everywhere, Arize, LangSmith, LangFuse, Helicone, Braintrust, Galileo, Monte Carlo, and incumbents like New Relic all racing to add agent-native monitoring. The investment thesis was simple: fight AI with AI, use agents to monitor agents.
The sharpest insight from the floor: building an agent takes two to three weeks; operating one takes months of continuous measurement. "The vendors pitching 'build agents in hours' are selling the first mile," one attendee noted. "The vendors talking about observability and Day 2 operations are selling the other 99%."
But observability only goes halfway. It answers what did the agent do? It does not answer the harder question: was the data the agent acted on trustworthy, and can I prove it?
The Grantex "State of AI Agent Security 2026" report put hard numbers on the gap. Of 30 major agent frameworks audited: not a single project assigns a cryptographically verifiable identity to each agent instance; 87% have zero action logging; 93% rely on unscoped API keys with no mechanism to restrict which operations the agent can perform, on whose behalf, or when access expires. Not one project produces an audit trail linking a specific action to a specific agent, authorization, and scope.
This is the accountability gap the conference kept circling without landing on. Observability tells you what the agent did. Verifiability tells you whether the data it consumed was intact and traceable. Those are different problems that require different infrastructure, and right now, only the first one has a vendor ecosystem behind it.
The ROI data makes this concrete. Deloitte's 2026 State of AI report shows 66% of organizations report efficiency gains, yet only 20% have seen revenue growth from AI. The gap is a trust problem. Enterprises will not scale what they cannot audit, and they cannot audit what their storage layer does not guarantee.
The Apache Iceberg Signal: Open Formats Won. Now What's Underneath?
Across town at the Marriott Marquis, the Apache Iceberg Summit told a complementary story. Last year's sold-out event drew nearly 500 attendees; this year expanded to two full days with speakers from Apple, Bloomberg, Pinterest, and Wells Fargo, and sponsorship from Databricks, Snowflake, Redpanda, and Ryft. The format wars are over, Hive is universally acknowledged as dead, outmatched on performance and portability. Iceberg won.
The technical headline was Apache Iceberg v3, now in public preview on Databricks: native geospatial data types, geometry and geography columns, enabling bounding box queries directly at the table format level, replacing the brittle hack of stuffing geodata into binary columns. Petabyte-scale deployments at companies like LinkedIn underscore how far the adoption has run.
Object storage has cemented its place as the persistent layer beneath the lakehouse. AWS recognized this with S3 Tables, a purpose-built Iceberg-optimized tier, but the all-in cost runs roughly $35/TB/month once compaction and monitoring fees stack up, and teams have documented costs running 20× higher than expected on high-frequency write workloads.
The open format thesis has a corollary that matters for the agent conversation: Iceberg solves the portability problem. When an agent pulls data from an Iceberg table, Iceberg tells you the schema, partitioning, and snapshot history. It cannot tell you whether the underlying data objects were tampered with between write and read, nor can it anchor that integrity proof to an independent, immutable ledger. That is the missing layer.
The Neocloud Explosion and Its Storage Gap
The neocloud narrative was impossible to ignore. Forrester projects neocloud revenue hits $20 billion in 2026. CoreWeave reached $5.1 billion in annual revenue in 2025, up from $16 million in 2022. Microsoft has committed over $60 billion in multi-year capacity agreements across Nscale, Nebius, CoreWeave, Iren, and Lambda.
The business model works for compute, Uptime Institute analysis shows neocloud GPU instances at roughly one-third the cost of hyperscaler equivalents. But the neocloud stack largely punts on storage. NVIDIA's BlueField 4 STX accelerates checkpoint I/O; VAST Data and WEKA provide the performance layer. The ecosystem is solving speed and cost.
None of it solves trust. When a neocloud customer runs an agentic workflow that checkpoints model state, retrieves external data, and writes a decision log, there is no cryptographic guarantee that any of those objects are intact and untampered. The neocloud stack has a verifiability gap at its foundation.
Data as a Moat, But Only If It's Verifiable
The "data as a moat" thesis ran through both conferences. The voice AI market crossed $22 billion in 2026, with 67% of Fortune 500 companies running production voice AI. The differentiator is not the model, it is the data. Voices.com's moat is consent-based licensing and voice provenance: the ethical origin and rights chain of every voice sample. The data is defensible because its lineage is verifiable.
The same pattern applies across every vertical. Companies that will win in the agentic era are not the ones with the best models. They are the ones with the most trustworthy data, cryptographically sealed, provenance-tracked, independently auditable.
One of the most telling observations from the week: the agentic AI teams and the data lake teams are operating in separate organizational worlds. One moves fast and ships agents. The other protects the crown jewels. When an agent makes a bad decision based on stale or corrupted data, the audit trail breaks at the seam between those two worlds. No one owns the full picture.
How We Think About This at Akave
Akave assigns a content identifier (CID) to every object at write time, creating a tamper-evident, cryptographically verifiable record anchored to an on-chain audit trail on Avalanche's L1 infrastructure. When an agent stores a checkpoint, retrieves a dataset, or logs a decision, the data's integrity is provable, not because the application logged it, but because the storage infrastructure guarantees it independently of the application layer.
This is the distinction HumanX conversations kept circling without landing on. Observability platforms can tell you an agent called a tool at 14:32:07 and received a response. They cannot tell you whether the data behind that response was the same data present when the dataset was approved by a human. Storage-level verifiability closes that gap.
Akave is fully S3-compatible. Existing Iceberg deployments, neocloud pipelines, and agentic workflows connect through a configuration change, swapping the endpoint and credentials, without rewriting application logic. The verifiability layer sits beneath the API surface, invisible to the application but auditable to anyone who needs it.
The agentic era does not just need faster or cheaper storage. It needs storage that answers the question Matt Garman posed at HumanX: how do you let autonomous systems take on real responsibility? You start by making their data verifiable.
What Comes Next
HumanX and the Iceberg Summit together painted an industry that has its compute story figured out, its data portability story emerging, and its agent deployment story accelerating. What is missing is the connective tissue: a verification layer that bridges the data lake teams and the agentic teams, that provides the trust substrate the neocloud stack lacks, and that gives regulators the audit surface they are about to demand.
Gartner has already flagged that over 40% of agentic AI projects will be canceled by end of 2027 if governance frameworks do not catch up. The companies that solve this will not just be compliant, they will hold the data moat that actually holds.
Get Started with Akave’s Object Storage for ai
Explore Akave's verifiable storage architecture and run a free trial at akave.com/free-trial.
Review how Akave integrates with AI and ML pipelines at akave.com/ai-ml-workloads, or connect via S3-compatible endpoint using the documentation at docs.akave.xyz.
FAQ
What was the key theme at HumanX 2026? The central theme at HumanX 2026 was the operationalization of agentic AI, the shift from experimentation and piloting to production deployment at enterprise scale. AWS CEO Matt Garman's keynote argued that the real barrier to agentic impact is not technology but how work, accountability, and decision-making are structured. The conference featured a dedicated Agentic AI Pavilion and sessions specifically focused on governance, production deployment, and what it takes to scale agents beyond the first mile of development.
What is agent observability, and why isn't it enough for agentic AI governance? Agent observability refers to monitoring systems that track what an AI agent does, which tools it calls, what outputs it produces, how latency and error rates behave. Tools like LangSmith, LangFuse, Arize, and Helicone represent the current state of this market. Observability answers "what did the agent do?", but it does not answer whether the data the agent acted on was intact, untampered, and traceable to its authoritative source. Verifiability is a separate layer that requires infrastructure at the storage level, not the application level.
What did the Grantex State of AI Agent Security 2026 report find? The Grantex report audited 30 major AI agent frameworks and found fundamental accountability gaps: not a single framework assigns a unique, cryptographically verifiable identity to each agent instance; 87% have zero action logging; 93% rely on unscoped API keys with no access controls tied to specific agents or authorizations. No framework reviewed produces an audit trail that links a specific action to a specific agent, a specific user authorization, and a specific set of permission scopes. The infrastructure for agent accountability, at the framework level, does not yet exist.
What were the major announcements at Apache Iceberg Summit 2026? The Iceberg Summit, held April 8–9 at the Marriott Marquis in San Francisco, featured speakers from Apple, Bloomberg, Pinterest, Wells Fargo, Databricks, Snowflake, and others. The biggest technical news was Apache Iceberg v3 entering public preview on Databricks, which introduces native geospatial data types (geometry and geography columns with bounding box predicate support), deletion vectors, row IDs, and variant data types. The format war was declared over: Hive is effectively obsolete, and Iceberg has become the de facto standard table format for lakehouse deployments.
How does Akave Cloud address the verifiability gap in agentic AI infrastructure? Akave assigns a content identifier (CID) to every object at write time, creating a cryptographically verifiable, tamper-evident record anchored to an on-chain audit trail on Avalanche's L1 infrastructure. When an AI agent stores a decision log, retrieves a dataset, or checkpoints model state, the integrity of that data is provable independently of the application layer, meaning it cannot be retroactively altered without breaking the cryptographic proof. Akave is S3-compatible, so existing pipelines can connect through a configuration change without rewriting application logic.
If observability tools already log agent actions, why do we also need storage-level verification? Observability tools log what the agent reported doing, the tool calls, outputs, and latency measurements that flow through the application layer. They cannot verify whether the underlying data that informed those actions was the same data that existed when it was approved, signed off, or validated. An agent could read a dataset that was silently modified after governance review, and an observability platform would have no way to detect it. Storage-level verification provides the root of trust that application-layer logging cannot: an independent, cryptographic proof that the data object at read time matches the data object at write time.
Isn't verifiability a problem for the agent framework to solve, not the storage layer? Agent frameworks can log actions and enforce policies at the orchestration level, but they depend on the data they receive being trustworthy. If the underlying storage is mutable and unverified, then framework-level controls are enforcing policies against data whose integrity they cannot independently confirm. The storage layer is the most defensible place to establish a root of trust, because it is independent of the application stack, the model, and the framework. A tamper-evident record at the storage layer survives framework changes, model updates, and organizational reorgs, it is the audit surface that remains when everything else changes.
What is the practical risk of the organizational silo between agentic teams and data lake teams? Most enterprises at HumanX described a world where the teams building agents (focused on orchestration and LLM integration) and the teams managing the data those agents consume (focused on governance and feature engineering) operate under different leaders, tools, and governance frameworks. The practical risk: when an agent makes a bad decision based on stale or corrupted data, the audit trail breaks at the boundary between these two organizational units. Neither team owns the full picture. Closing this silo requires a storage substrate that provides verifiable provenance regardless of which team, tool, or compute engine touches the data.
Further Reading
- The Neocloud Storage Problem: Why GPU-Rich Clouds Still Need a Persistent Object Store, the structural gap in neocloud infrastructure that HumanX confirmed on the expo floor
- Cloud Egress Fees Explained: Azure vs Google Cloud vs Akave, the cost math behind the egress multiplier that compound against AI workloads on hyperscalers
- The FinOps Blind Spot: Cloud Waste You Can't See, why storage costs grow invisibly during AI scale-up