AI in the News

AI in the news: week of March 22, 2026

GTC 2026 anchors the week. Vera Rubin, a $1T order book through 2027, and a partnership map from Uber to Disney to Eli Lilly. Mistral ships Small 4 and announces Forge for training-on-your-own-data. AI hits 25% of March layoffs. The EU's child-safety amendment lands.

Sid Smith

22 Mar 2026 • 6 min read

What this week actually changed: NVIDIA consolidated around platform-not-just-chips with a $1T order book, Mistral made a credible "bring the model to the data" bet against the hosted-AI default, and AI hit 25% of US March layoff announcements, the highest single-cause share in the history of the Challenger report.

GTC week. Jensen Huang on stage Monday March 16, the rest of the industry calibrating against the announcements for the next four days. Vera Rubin is the headline; the order-book number is the headline-of-the-headline; the partnership map is the part with the longest tail. Mistral used the same week to ship Small 4 and announce Forge. The Challenger report for March landed midweek. The EU added a CSAM provision to the AI Act.

NVIDIA GTC 2026: Vera Rubin, plus a $1T order book

March 16, San Jose. Huang delivered the GTC 2026 keynote to a sold-out arena and used it to launch the Vera Rubin platform, a new Vera CPU paired with the Rubin GPU, 288 GB of HBM4, 50 PFLOPS of FP4 compute, and a claimed 10x inference-per-watt over Blackwell. AWS, Google Cloud, Microsoft, and OCI are the first hyperscalers in the deployment queue, with CoreWeave, Lambda, Nebius, and Nscale on the NVIDIA Cloud Partner side.

The number that traveled was the order-book one. Huang said on stage that he expects $1 trillion in orders for Blackwell and Vera Rubin systems through 2027, double the $500B figure from a year ago. Treat that number as the forward-looking statement it is. The internal logic is: hyperscaler capex commitments plus sovereign-AI build-outs plus the agent workloads that haven't shipped yet add up to a number that justifies the build-out NVIDIA is already in the middle of. Whether the demand lines up with the supply is the question the next four quarters of earnings will answer.

The partnership announcements are the part I'd watch over the model specs. Uber will deploy NVIDIA Drive AV across 28 cities and four continents by 2028, starting in LA and SF next year. Disney showed an Omniverse-trained robot. Eli Lilly is on the healthcare-and-pharma roster. The pattern is NVIDIA-as-platform-foundation across verticals, not NVIDIA-as-chip-vendor, and that's a meaningfully different competitive surface. The chip business sells silicon. The platform business sells the developer-and-partner stack that runs on the silicon. The latter is much harder to displace.

The bit I'd flag for the principled-practitioner crowd is the OpenClaw / NemoClaw framework Huang positioned as "an open-source agentic AI framework, like Linux." If that framing actually holds (open-source, vendor-neutral, runnable on non-NVIDIA hardware) it cuts against the lock-in the rest of the keynote was building. If it ends up being open-in-name with NVIDIA-specific dependencies, it doesn't. I'll be reading the license and the reference implementation, not the keynote slide.

Mistral week: Small 4, Forge, and bringing the model to the data

March 16, same day as the GTC keynote. Mistral released Mistral Small 4, unifying the previously separate Magistral (reasoning), Pixtral (multimodal), and Devstral (coding) lines into a single model. They also dropped Leanstral, a Lean 4 formal-proof agent, the same day. March 17, at GTC, they announced Forge, an enterprise platform for training frontier-grade models on proprietary data. Voxtral TTS landed March 23 to round out the run.

The unification move on Small 4 is the strategic part. Three product lines collapsing into one is what you do when the model is good enough that the specialization tax isn't paying for itself. It's also what you do when the engineering org needs the maintenance burden cut. Either reading is a sign Mistral is operating from a stronger position than the "European underdog" framing usually implies.

Forge is the more interesting announcement to me. The pitch is that an enterprise can use Forge to train its own frontier-grade model on its own proprietary data, on its own terms, without the data leaving its boundary. If that holds in practice (and the early documentation suggests it does) Forge is a direct response to the data-gravity problem I keep writing about. Sensitive data doesn't have to go to the model if the model can come to the data. Mistral is betting the European compliance posture plus the on-premises training story is a viable wedge against the hyperscaler-default. I think they're right that it is.

The competitive question is whether Forge actually ships at the quality the announcement implies, and whether the price tag puts it in reach of the mid-market or only the very largest enterprises. I'll watch the first three customer references.

The EU adds a CSAM provision to the AI Act

March 13, just inside the prior week, the Council agreed a position to streamline parts of the AI Act and added a new provision prohibiting AI generation of non-consensual sexual or intimate content and CSAM. The amendment is straightforward in its substance and overdue in its timing. I'd put it in the same category as SB 53 from last October, narrow, targeted, achievable, the kind of regulation that closes a specific harm rather than trying to write the whole framework at once.

The streamlining piece is the part that needs more scrutiny. The same Council session moved the application date for high-risk AI rules and softened the obligations for some GPAI tiers. The framing is "simplification"; the practical effect is that some of the compliance machinery that was supposed to bind in 2026 now binds in 2027 or 2028. The lobby pressure to slow the timeline has been heavy and consistent and is clearly working at the margins. Whether that's the right tradeoff depends on whether the extra year produces better-engineered compliance or just more time to find loopholes.

What I'll watch: how the simplified GPAI obligations look in the final text once the Omnibus VII package lands in the spring, and whether the CSAM amendment gets bolted onto US state-level frameworks in the next legislative cycle.

AI at 25% of US March layoff announcements

Challenger's March report puts AI at 25% of announced US cuts, 15,341 of 60,620 (Challenger). The trend continues. Oracle's reported 20,000-30,000 reduction is framed as freeing budget for AI capex, not as automation displacement, which is the pace problem I keep flagging. Longer position here.

A few smaller items worth flagging

GPT-5.4 mini and nano, OpenAI's smaller-tier additions to the GPT-5.4 family released earlier in March, shipped on March 17 and absorbed into the OpenAI API the same day. The pricing positions them squarely against Gemini 3.1 Flash-Lite, a tier-vs-tier comparison the buying side should be running.
MCP crossed 97 million installs by end of March per Anthropic's quarterly update, a milestone that puts it firmly past the "experimental standard" framing and into infrastructure territory. MCP (this is the Model Context Protocol, a way for AI models to talk to tools and data sources in a standard format, if you want to look it up later) has become load-bearing for the agent stack in a way nobody quite predicted a year ago.
GE 26 Benchmark Wargame ran March 13-27 in Alexandria with the Air Force using its WarMatrix AI environment for the first operational time. Worth flagging as the kind of defense-AI procurement that doesn't make the consumer headlines but shapes the budget conversation in DC.

What to watch next week

The infrastructure layer is consolidating around NVIDIA-as-platform. Vera Rubin is the chip headline; the partnership-and-platform map is the moat. The principled-user response is the same as it's always been: build the agent stack so it isn't structurally bound to a single vendor, even when that vendor is the one shipping the best hardware. OpenClaw being actually-open or not-actually-open is the test case to watch.

Mistral's Forge is the most interesting bet against the hosted-AI default. Bringing the model to the data instead of sending the data to the model is the right architectural answer for a long list of compliance-bound enterprises. If the execution holds, this is the European bet that pays off.

The labor numbers are the constant. AI hit 25% of March layoff announcements. The pace is wrong, the framing is convenient for the cutters, and Oracle's capex-funding framing is the cleanest version of the pace problem.

Next Sunday: GTC after-effects (the press cycle continues for a week), DeepSeek V4 rumors getting louder, whatever lands midweek.