AI4 2026 Recap - From AI Experiments to AI Workloads
In this blog
In this blog
- The scale of AI4 says something on its own
- From projects to workloads
- Speaker and workshop themes
- What we heard on the floor
- Where the practical content lived
- Where WWT and Softchoice fit
- Looking ahead
The scale of AI4 says something on its own
AI4 2026 ran August 4–6 at The Venetian in Las Vegas, its ninth year. Organizers reported more than 12,000 attendees from 100 countries, up from 8,000 the year before. Exhibitors and sponsors grew from 250 to over 400. Startup Alley went from 25 companies to 66. Credentialed media more than doubled, to 225.
Read those numbers together and you get the shape of the market: roughly 50 percent more buyers walking the floor, against a near-doubling of vendors competing for their attention. That imbalance turns out to explain a lot about the conversations we had.
The program ran more than 1,000 speakers across some 80 education tracks: AI transformation, industry verticals, job functions, special interests and deep technical content. Financial services, healthcare, retail, manufacturing, energy, transportation, insurance, aerospace and defense, government and public sector, each with its own agenda.
That breadth matters because AI4 is vendor-neutral. Nobody anchors the program to a single provider's roadmap. When a conference like that grows 50 percent in a year, it tells you AI has stopped being a platform story and become an operations story. Our team came away impressed by the targeted content and the caliber of the speakers, and struck by how quickly this event is maturing.
From projects to workloads
Here's the shift we kept hearing, in the vocabulary people used without noticing.
A project is scoped, funded once and judged on whether it shipped. A workload runs continuously. It consumes capacity, carries a unit cost, needs service levels and someone owns it at 2 a.m. when it misbehaves. AI4 2026 was the year enterprise AI conversations switched from the first word to the second.
Six operational concerns came up repeatedly, in sessions and in hallway conversations alike:
Capacity planning came back. Inference is a sustained load now, not a burst. Where it runs (public cloud, colocation, on-premises, edge) is a real architectural decision again, driven by data gravity, latency and sovereignty.
Token cost became a line item. Teams that shipped in 2025 are being asked what each transaction costs. Model routing, caching, distillation and small-model substitution came up constantly. AI has a unit economics problem, and finance has noticed.
Evaluation matured into a discipline. Golden datasets. Regression suites for prompts. Offline versus online evaluation. The teams furthest along treat evals the way good engineering teams treat test coverage.
Observability became non-negotiable. Not logging model calls: tracing a decision across a multi-step agent run, through every tool it touched.
Security and governance stopped being someone else's track. Prompt injection, tool-permission scope, data leakage through context windows, agent identity. These showed up inside architecture sessions, not just in the security hall.
The guidance gap widened. This one surfaced less from the stage than from the people standing in front of it. Companies arrived with a clear sense of what they want AI to accomplish and far less clarity on how to begin. The hesitation wasn't about ambition or budget. It was about taking a first step in a fast-moving field without someone experienced alongside them, and more than a few described programs sitting still not because a business case failed, but because nobody wanted to own the first architectural decision alone. Of everything we heard across three days, this was the most consistent: the appetite for AI transformation has outrun most organizations' confidence to start it.
Speaker and workshop themes
With 80 tracks running, nobody saw all of AI4. But five themes came through as common feedback across the program, and they map closely to what we heard at the booth.
1. Treat agents as labor, not software
The opening keynote was The One Decision That Separates AI Winners from AI Casualties," and it resonated widely with people running AI programs.
The session led with survey findings: 96 percent of enterprise leaders believe their employees are using unsanctioned generative AI tools, 80 percent of CIOs say their role is at risk without measurable AI ROI, and 77 percent expect a peer to be fired over a failed AI strategy. Vendor research, and worth reading as such.
The sharper point was the gap between perception and reality. Many CEOs estimate their organization runs around 40 AI agents; the research suggests large enterprises are often closer to 400, most built independently by employees looking for a productivity win. The speakers described these as vibe-coded agents and drew a parallel to spreadsheets spreading through corporations 40 years ago, well before IT understood the implications.
The recommendation was direct. IT should move out of the gatekeeper role and into something closer to HR for agents, managing identity, access and governance, which depends on having visibility into what is actually being built. And if agents are being used as labor, the argument went, they should be treated as labor: placed on the P&L, with governance expressed as a budget line carrying someone's name.
2. AI is an infrastructure story now
"The AI Reckoning: Chips, Constraints and the Next Generation of Compute" was another popular session.
The framing was that demand for chips and memory has become insatiable (the oil of AI) and that the economics need to improve substantially for the technology to capture the opportunity in front of it. The through-line was that future breakthroughs will depend as much on semiconductors, networking, electrical power and capital as on algorithms.
The same ground got covered from the buyer's side elsewhere in the program. The presenters talked about the infrastructure demands created by the shift from assistants to autonomous agents and there were points made on the case for purpose-built AI over rented capacity.
If compute cost and energy availability bound what is worth automating, the differentiator moves toward direction and verification: deciding what to point the capability at and confirming it did what it reported.
3. Openness, regulation and who sets the pace
The headline session, "Architects of Intelligence," was the keynote session on stage Wednesday to a standing-room-only theater.
The presenters disagreed productively. One presenter suggested AI could surpass human capability across nearly every domain within 20 years and likely sooner, and pointed to white-collar displacement in areas such as call centers and paralegal work. Another presenter countered that AI redefines tasks rather than eliminating roles, and argued that large incumbents tend to overstate the threat in ways that conveniently limit competition. The third presenter reframed the debate around a different measure, cautioning that increased productivity does not automatically translate into shared prosperity.
Where they converged was more striking. All three supported regulation and made the case for keeping AI open, sharing a concern about a small number of companies controlling the pace of progress. Hinton's framing was that while no industry welcomes regulation, this one should want it, because the objective is for AI to help people.
4. Sovereignty and open-weight models
Sovereignty ran through the program as a distinct theme: organizations wanting to own their AI, their data and their infrastructure rather than handing everything to a vendor.
The session, "Pioneering Transparent and Transformative Frontier AI," made the technical case, outlining a family of open-weight models organizations can deploy inside their own environments rather than consuming through closed APIs. For enterprise teams, the calculus is straightforward: on-premises deployment offers more control over sensitive data while containing token costs that scale quickly once hundreds of agents are running.
5. Responsible AI moved from principle to practice
"AI's Race to Reinvent Medicine," was an interesting session that showed AI well past demonstration and into real therapeutic pipelines: accelerated drug discovery, disease modeling and longevity research. The panel also made the case that governance needs to mature alongside capability, given how broadly these same techniques can be applied.
Two presenters made an adjacent argument in "AI is Not Value-Neutral. Now What." session. The useful distinction: analysis tells you what is and what is likely, while deciding what to do about it is a normative act that stays a human one no matter how automated the analysis becomes. Pricing tiers, eligibility rules and exclusion criteria are often treated as analytics output. Her point was that each should have a named owner before an agent executes against it at volume.
What we heard on the floor
World Wide Technology and Softchoice shared a booth at AI4 exhibit hall, and it ran on one line:
-WWT + Softchoice booth tagline, AI4 2026
Six words doing the work of a thesis statement. Paired with a joint WWT and Softchoice presentation and swag worth stopping for, it gave people a reason to slow down in a very loud room. Those conversations turned into more than 15 booked meetings with prospective customers and partners over three days.
The alternative it points to is concrete: validating an AI stack against real infrastructure, on AWS, before committing to it. Real data, not vendor promises.
Three questions came up again and again.
"Can you help us navigate the AI ISV landscape?" This was the most common request we fielded. It says something about where the market is: in a hall with more than 400 AI exhibitors, the most valuable thing on offer was help working out which of them fit a given environment. The ecosystem is rich with capable options. What buyers wanted was a way to narrow it with confidence.
"How do we get from idea to production?" People responding to the tagline itself, which tells us the framing landed. Closing the distance between a working demo and a running workload is where most of the energy is right now.
"What does WWT and Softchoice together mean for us?" Softchoice, now a World Wide Technology company, serves the SMB and commercial segment, while WWT focuses on large enterprise. Together that spans the full range of customer size, with a portfolio across software, cloud, cybersecurity and AI, and a shared AWS practice underneath it. The ISV question came from organizations of every size, and the right answer looks different at 200 employees than at 200,000.
The overall tone was constructive. Less concern about being left behind, more focus on building things that will hold up in production: on cost, on security, on operations. The questions have gotten more specific, which is a sign of a market maturing rather than cooling.
Where the practical content lived
A useful note for anyone planning a 2027 agenda.
The keynote theater was at its best on the big questions: where the technology is heading, what it means for work, who should set the pace. That content was genuinely strong, and it is the reason this event has grown the way it has.
The applied, how-we-actually-built-it material tended to live in the track rooms and along the exhibit hall aisles. Both are worth the time, but they serve different purposes. Teams sending people to AI4 specifically to bring home implementable detail will get more out of weighting their schedule toward the tracks and the floor, and treating the keynotes as the strategic context around them.
That pattern also explains the question we heard most at our booth. With more than 400 vendors presenting answers, information was not the scarce resource. A way to work out which answer fits a particular environment was.
Where WWT and Softchoice fit
The capability we showed at AI4 speaks directly to the ISV and production-readiness questions we heard most.
The AI Proving Ground is a production-grade lab built on WWT's Advanced Technology Center, backed by WWT's infrastructure and delivered together with Softchoice and AWS. It holds real hardware from more than 200 OEMs. Customers can run their own AWS workloads on it, compare options side by side and validate security, performance, governance and cost before committing to production.
The joint model covers the full range. WWT brings the infrastructure and engineering depth. Softchoice, now a World Wide Technology company, brings reach across the SMB and commercial market along with its software and cloud practice. AWS brings the platform most of these workloads run on.
Alongside the lab, WWT works with customers on AI readiness and data foundations, platform and ISV selection, agentic reference architectures and the operating-model questions of who owns what and how decisions get approved. Most of that work starts as a briefing or a working session.
Looking ahead
AI4 returns to The Venetian August 3–5, 2027. Between now and then, the themes worth tracking are the ones that recurred all week: agents moving onto the P&L with named owners, infrastructure and energy as the real constraint on what gets automated, sovereignty and open-weight deployment, and a governance conversation that is getting steadily more specific.
If you were at AI4 and saw something we missed, we'd like to hear it. And if you're working through any of these questions inside your own environment, that's a conversation we're always glad to have.