AI in the News

AI in the news: week of March 1, 2026

MWC Barcelona pushes on-device AI back to the front of the stack, AMD and Nutanix sign a multi-year agentic-AI deal that makes the on-prem enterprise stack real, HyperNova 60B drops on Hugging Face, and the state-AI-law map keeps moving while the federal preemption clock ticks.

Sid Smith

01 Mar 2026 • 7 min read

What this week actually changed: the hardware layer pushed back to the front of the conversation. On-device AI is the MWC story; a real non-hyperscaler enterprise stack arrived in the form of an AMD-Nutanix deal; and the state-level AI law map kept building while the federal preemption clock counted down.

MWC Barcelona opens Monday March 2, and the device-and-network announcements that always cluster around it shaped most of the week's stories. Quieter on pure model releases than the back half of February. Louder on infrastructure and policy. Worth covering both ends.

MWC: on-device AI as the through-line

MWC Barcelona's Day 0 and Day 1 announcements landed on March 1 and 2. The framing the show is using this year is "the IQ era," and the Day 1 lineup leans hard into agentic AI in the device formfactor. Honor showed off what they're calling a "Robot Phone", agentic AI running locally on the handset, paired with a humanoid-robot stage demo. Samsung used the Galaxy S26 launch to push on-device Galaxy AI plus a choice of agents (Bixby, Gemini, Perplexity) running across phone and wearables. Oppo's Find X9 series has live translation and AI Portrait Glow running on-device. Xiaomi launched the Xiaomi 17 Ultra globally. Motorola detailed the Razr Fold on a Snapdragon 8 Gen 5. Samsung and NVIDIA also confirmed a multi-cell vRAN test combining Samsung's virtualized RAN with NVIDIA's accelerated compute, with the validation announcement timed to MWC.

The pattern across all of it is on-device AI as the differentiator and agentic features as the new category vocabulary. Samsung's pre-show announcement is the cleanest framing of the bet. Galaxy AI working across the device estate without needing a cloud round-trip for the high-frequency interactions. Honor's Robot Phone is the more aggressive bet, the phone-as-agent positioning collapses a lot of stack assumptions if it actually works.

I like this direction more than most of what's happened in the hosted-AI category in the last 18 months. On-device AI is structurally aligned with the principles I keep writing about: sensitive data stays on the device that owns it, the user keeps the option to operate without sending everything to a hosted backend, and the hardware vendor's incentive to make the on-device case compelling pushes the model-efficiency frontier in a direction the hyperscalers were never going to push it on their own. The MWC framing is finally a counterweight to the "AI = remote API" assumption that the last two years normalized.

The skepticism I'd hold: most of the on-device demos at MWC will quietly fall back to a hosted call for anything actually difficult, and the "agent on the phone" pitch is years from being load-bearing. Worth watching how much of what was announced this week is on-device in production versus on-device in the keynote. The category is real, the marketing is ahead of the engineering, and the engineering is catching up faster than I expected.

AMD and Nutanix make the on-prem case real

February 25, AMD and Nutanix announced a multi-year strategic partnership focused on agentic AI infrastructure for enterprise. The first jointly developed agentic AI platform is targeted for late 2026, with a broad set of OEM server partners signed up to ship the integrated stack.

Two things matter here. First, it's another non-NVIDIA, non-hyperscaler enterprise-AI bet getting real partner alignment. AMD is positioning Instinct accelerators plus EPYC plus the Nutanix hyperconverged platform as the on-prem agentic-AI foundation that doesn't require renting a hyperscaler. The OEM coverage matters here, enterprise buyers want a stack they can buy from Dell, HPE, Lenovo, and put in their own data center, and that's what this announcement enables. Second, the explicit "agentic AI" framing in an infrastructure deal is the signal that the agent stack is now the buying conversation. Two years ago this would have been an "AI training cluster" announcement. The shift to "platform for agents" is the same shift Anthropic and OpenAI made at the application layer, now landing in the infrastructure conversation.

This is the on-prem case gaining real vendor mass behind it. AMD plus Nutanix plus the OEM tier is enough commercial gravity to make the on-prem agent stack a default option for enterprise buyers in 2027, not an exotic one. I'm pleased to see this. The structural answer to data-sovereignty and lock-in concerns is a viable non-hyperscaler stack that enterprise procurement teams can actually purchase, and this deal is a concrete step toward it.

HyperNova 60B and the open-weights compression story

February 24, Multiverse Computing released HyperNova 60B 2602 as a free download on Hugging Face. The model is a 50% compressed version of OpenAI's gpt-oss-120B, with measurable improvements in tool calling and agentic coding compared to the prior HyperNova generation.

Worth flagging two things. First, the compression-as-research-direction story is finally producing models people actually want to run. Halving the parameter count of a frontier-grade open-weights model while improving agent-task performance is the kind of result that makes on-device and on-prem deployments much more practical for organizations that don't have hyperscaler budgets. Second, this is downstream work on OpenAI's gpt-oss release, which keeps the open-weights ecosystem productive in a way that pure-open-source frontier labs don't have to carry alone. The compounding is real.

For organizations with frugal AI budgets, 60B-class models that fit on a single high-end GPU and still handle agent workloads change the deployment math. The capability gap to hosted frontier is closing on the practical workloads, not on every benchmark, on the workloads enterprises actually run. That's the part that moves the buying conversation.

The state-AI-law map keeps moving

The state-by-state AI law map kept moving this week. Per the February 23 Troutman update, bills crossed chambers in four states (Oregon, Utah, Virginia, Washington), advanced out of committees in four more (Idaho, Iowa, Oklahoma, Hawaii), and were newly introduced in California, Colorado, Iowa, and Georgia. Washington's HB 2225 (companion-chatbot regulation) passed the House 69-28 on February 17 and got a do-pass vote out of the Senate Environment, Energy, and Technology Committee on February 20.

Meanwhile the federal preemption pressure keeps building. The White House executive order from earlier in the quarter directed the FTC to issue a policy statement by March 11 describing how the FTC Act applies to AI and when state laws "requiring alteration of truthful outputs" are preempted by federal law. Colorado postponed its AI law enforcement date from February 1 to June 30, 2026, after industry pushback, with a March working-group draft pointing toward a January 2027 reset. The Cooley state-law tracker is the cleanest reference for the current posture.

This is governance happening at the speed it actually happens. SB 53 in California set the precedent in late 2025; New York's RAISE Act extended it in early 2026; the rest of the states are now negotiating their versions in real time, and the federal preemption fight is the friction layer slowing the convergence. I keep saying governance is the work, and what I mean by that is the messy state-level negotiation we're watching, not the headline federal action that hasn't arrived. The preemption clock matters, if the FTC's March 11 statement is broad, it could blunt the state-law momentum that's been the actual driver of AI accountability since SB 53. If it's narrow, the state map keeps building and the federal layer eventually has to meet it. I'd bet on narrow, and I'd hedge that bet.

A few smaller items worth flagging

Block headcount and the Q1 trend: Q1 2026 tech-industry layoffs hit nearly 80,000, with roughly half attributed to AI by the trackers. Block's nearly-halved workforce is the headline number; I've worked the Dorsey framing and the pace-vs-direction read in the Feb 8, Feb 15, and Feb 22 roundups, and the longer position is in the job-security piece. Nothing new this week beyond the trend continuing.
Anthropic Series G: Anthropic disclosed that more than 500 enterprise customers are each spending over $1M annualized on Claude. The enterprise revenue concentration is the more interesting story than the funding round itself.
Snowflake-OpenAI partnership keeps expanding through February, the $200M multi-year deal announced earlier in the month is now landing in customer-facing integrations. Watch how the data-residency story plays out as more enterprises wire their warehouses to OpenAI endpoints.
EU AI Act: The Commission's February guidance on high-risk AI use cases hit on schedule, and the Chapter V GPAI obligations remain in effect. The full Act applicability date (August 2, 2026) is now five months out, and the compliance machinery is finally getting urgent.

What to watch next week

The hardware layer reasserting itself is the most welcome shift in two years. MWC's on-device AI push is the clearest counter-positioning to the hosted-frontier default we've seen. Phone vendors and chip vendors have a structural incentive to make local AI work, and they're spending against it. Good for the principles I write about. Watch the gap between keynote and shipping.

The non-hyperscaler enterprise stack is getting real. AMD plus Nutanix plus OEM coverage is a buying motion for on-prem agentic AI that didn't exist two years ago. Combined with open-weights compression work landing on Hugging Face, the practical alternative to "rent everything from a hyperscaler" is now something enterprise procurement can actually evaluate.

Next Sunday: MWC week-in-review with the actuals from the show floor, the FTC's March 11 statement on AI preemption, and whatever else lands in the lead-up.