AI in the News

AI in the news: week of March 8, 2026

MWC Barcelona puts agentic AI in every booth, GPT-5.4 ships mid-week, Oracle telegraphs the largest layoffs in its history to fund AI capex, and the EU Council adds non-consensual deepfakes to the AI Act's prohibited list. The week the cuts started being about funding compute, not replacing workers.

Sid Smith

08 Mar 2026 • 7 min read

What this week actually changed: the AI-cited layoff story split in two. The front-line "AI replaced the workers" cuts are still happening, and now Oracle is telegraphing tens of thousands more cuts framed openly as "we need cash to build inference capacity." Both shapes are happening at the same time.

MWC Barcelona was the centerpiece of the week. The show ran March 2 to 5 under the "IQ Era" banner, and almost every keynote and booth pitch led with agentic AI rather than 5G or device specs. Mid-week, OpenAI shipped GPT-5.4, Oracle telegraphed what looks like the largest layoff in the company's history, and the EU Council folded a new prohibition into the AI Act.

MWC: agentic AI as the product layer

MWC Barcelona 2026 opened March 2 with the "IQ Era" theme, and the operative word in nearly every keynote was "agentic." Qualcomm's Cristiano Amon framed his keynote around 2026 as the year agents redefine the digital experience past the smartphone, unveiling AI-native modems, RAN, and Wi-Fi platforms plus a Snapdragon Wear Elite chip targeted at AI pendants and display-free smart glasses. Deutsche Telekom debuted the Magenta AI Call Assistant, built with ElevenLabs and running inside the call network rather than as a phone app, plus MINDR, a multi-agent network operations system built on Gemini that monitors and corrects performance across radio, transport, and core layers autonomously.

Two things stood out structurally. First, agents at the network layer, not just at the device or app layer. Mavenir and Mycom announced agent-to-agent integration between their network-assurance copilots; NVIDIA partnered with several large operators to deploy specialized AI agents for customer care and network operations and to scale "sovereign AI factories" for enterprise demand. Second, AI-RAN moved from concept slide to live deployments. Ericsson and Intel announced they'll collaborate on AI-native 6G, with commercial 6G targeted for 2029.

The sovereign-AI-factory framing matters. Telcos pitching themselves as the operators of regional AI infrastructure is the same distribution-vs-concentration story I keep returning to. The hyperscalers want every workload on their capacity. The telcos want a slice of inference at the network edge, in-country, on operator-controlled hardware. That's structurally the right direction for the frugal-AI strategy, more places to put inference, more leverage on the price-and-residency conversation, less of the "your data goes to one of three US companies or it doesn't move at all" architecture.

The thing I'd push back on at MWC is the call-network voice-assistant pattern. Magenta AI Call Assistant is a useful product. It is also a layer of inference sitting in the middle of every customer phone call, with whatever the speech-to-text and LLM pipeline retains. The pitch is convenience; the underlying transaction includes routing all your call audio through a hosted AI stack. The PII implications aren't new (telcos already have the call records) but the inference layer is a new processing stage with new retention, new training-data exposure, and new vendor dependencies. Worth tracking what the data-handling commitments actually are once the marketing settles.

GPT-5.4 ships, and the labs are on a 28-day clock

OpenAI released GPT-5.4 on March 5. The release came in three variants and posted record scores on computer-use benchmarks plus material reasoning gains over GPT-5.3. Some press framing called this a rushed release in response to the Sonnet 4.6 and Gemini 3.1 Pro launches the prior fortnight; the cadence does fit that read, but the model itself is real.

Two notes. The computer-use benchmark gains are the headline, and the headline matters because computer-use is the capability that turns "model in a chat window" into "agent that can run your browser." Anthropic put real distance into computer-use with Sonnet 4.6 in February; GPT-5.4 closes some of that gap. The frontier labs are converging on the agent-driving-the-OS use case and the benchmark race is now squarely there.

The other note is the release tempo. Anthropic shipped Opus 4.6 on Feb 5, Sonnet 4.6 on Feb 17, Google shipped Gemini 3.1 Pro on Feb 19, and OpenAI is now on GPT-5.4 by March 5. Four major frontier releases in 28 days. The labs are on a near-monthly clock and the deltas between releases are narrower than they were a year ago. The cumulative quarterly delta is still large, and the practical answer for anyone building on this stack is to keep your wrapper layer thin enough that the underlying model can be swapped without a rewrite. Same point I made when Sonnet 4.6 landed, and it gets more true every release cycle.

Oracle telegraphs the largest layoffs in its history, to fund capex

Bloomberg reported on March 5 that Oracle is preparing what analysts at TD Cowen estimate will be 20,000 to 30,000 cuts (roughly 18% of Oracle's 162,000 global workforce) to free up an estimated $8 to $10 billion in annual cash flow for AI infrastructure capex. Execution started at the end of the month per follow-up CNBC reporting, with affected employees receiving termination emails from "Oracle Leadership" with no manager warning. The capex commitment driving the cut is roughly $156 billion over the buildout horizon.

The financial framing here is important and worth being precise about. Oracle posted a 95% jump in net income last quarter to $6.13 billion, with remaining performance obligations at $523 billion (up 433% year over year). This is not a company in revenue distress. It is a company making a capital-intensive bet on AI infrastructure that its current balance sheet can't comfortably sustain, and eliminating tens of thousands of employees to close the gap.

I want to be plain about where I sit on this. The displacement is real and it's accelerating faster than I expected. The Oracle cut isn't the front-line "AI replaced the workers" story. It's the back-end "we need cash to build inference capacity, and headcount is where the cash comes from" story. Both shapes of the cut are happening at the same time. The pace is what's wrong, and the pace is driven by short-term incentives. Markets reward the cuts on this earnings call. They don't price the rebuild cost two earnings cycles out, and they don't price the operational gaps that show up when the institutional knowledge walks out the door at 6am with no notice.

The sustainable working model is human+AI collaboration. The firms that figure out the collaboration will outperform the firms that just cut. To be clear: the headcount still shrinks under collaboration. It just shrinks less and shrinks well. The other enterprise software companies on the same capex/headcount squeeze (Salesforce, Workday, ServiceNow) are watching this closely. Expect copycat restructurings through the rest of 2026. I'd rather be wrong about the pace than be caught flat-footed. See the longer essay on the labor question for the full position.

The EU adds a deepfake prohibition to the AI Act

The Council of the EU agreed its position on the AI Act streamlining package this week. Two pieces matter. First, a new prohibition was added for AI practices that generate non-consensual sexual or intimate content, or child sexual abuse material, pulling those use cases into the Article 5 prohibited-practices list directly. Second, the high-risk timeline shifted: stand-alone high-risk AI systems now have a December 2, 2027 application date and embedded high-risk systems push to August 2, 2028.

The deepfake prohibition is the right addition and should be celebrated as such. The technology to generate non-consensual intimate imagery has been runnable on consumer hardware for two years. The Sora 2 character feature I flagged in October put a frontier-lab product in the same category in a thinner wrapper. Pulling these use cases into Article 5 (the no-mitigation-can-make-this-okay tier) is the regulatory shape that actually matches the harm. It's also the kind of carveout that translates cleanly into other jurisdictions, which matters because state-by-state US legislation on this has been patchy.

The high-risk timeline shift is the more contentious move. Pushing stand-alone high-risk AI to December 2027 is roughly a 16-month delay from the original schedule, and the embedded slice goes longer. I think the underlying instinct is right, the conformity-assessment infrastructure isn't real at the scale the original timeline assumed, and theatrical compliance against a deadline the auditors can't actually enforce is worse than an honest delay. But the criticism worth taking seriously is regulatory drift: every push of the timeline is a window for the obligations themselves to get diluted. The governance work itself doesn't change shape because of the calendar; what changes is the political pressure on what compliant ends up meaning by the time the deadline arrives.

A few smaller items worth flagging

Apple unveiled the MacBook Neo at $599 with the A18 Pro chip, positioned as a mass-market on-device AI laptop. The device-side AI story is starting to move, sub-$700 hardware that runs reasonable on-device inference shifts the build-vs-buy math for a lot of small-org workflows.
Mistral Small 4 shipped in early March under Apache 2.0, continuing the Mistral licensing shift I flagged in the Feb 22 roundup. Frontier-tier open weights from a European lab is exactly the foundation the individual-cognition-as-IP and on-prem arguments need.
NVIDIA's MWC announcements with telco partners put specialized AI agents into customer-care and network-ops workflows at multiple operators. Worth watching how the agent stack is structured, whether the operator runs the inference or NVIDIA does is the load-bearing detail.
DeepSeek V4 preview chatter kept building during the week, with the actual release landing later in the quarter. The China-side cadence pause for Lunar New Year is over and the next wave is queued.

What to watch next week

Agentic AI is now the product layer the industry is building toward. MWC made it explicit: the device, the modem, the call network, the RAN, all of it is being repositioned as foundation for autonomous agents. This isn't a research conversation anymore; it's a deployment conversation, and the architecture decisions the operators make this year about where the inference runs will shape who controls the agent layer for the next decade. Distribution beats concentration here because the alternative is a handful of US hyperscalers running every agent on the planet.

The labor curve steepened again. Oracle's 30,000-cut signal is the largest single AI-driven restructuring of the cycle so far, and the framing is explicit: the cuts fund the capex. Plan for the realistic view; hope for the slower one.

Next Sunday: GPT-5.4 in the wild after a week of developer testing, MWC follow-throughs as the booth promises hit reality, and whatever lands midweek.