> daily_signal(2026_05_08)

Same week, opposite directions: the EU loosened its AI rulebook while the US tightened its own.

PickBits Daily Signal · Friday, May 8, 2026

By Mark Pickering · 6 min read · May 8, 2026

// tl;dr

The EU pushed its high-risk AI rules out 18 months. Council and Parliament struck a provisional deal May 7. The bulk of high-risk obligations now apply December 2, 2027 (general) and August 2, 2028 (safety components in regulated products). The transparency grace period gets cut from six months to three.
Three more public layoff announcements that name AI. Cloudflare cut 1,100 (about 20 percent of staff), pinning it on internal AI use up more than 600 percent in three months. BILL announced up to 30 percent headcount and a $30 to $60 million restructuring charge. Upwork cut about 145 (24 percent). All three landed within 24 hours.
NVIDIA and IREN added 5 more gigawatts to the AI build-out. A May 7 partnership puts NVIDIA's DSX-aligned infrastructure on IREN sites globally, with a 2-gigawatt Texas campus as the flagship. NVIDIA gets a 5-year right to buy up to 30 million IREN shares at a $70 strike, total potential investment up to $2.1 billion.
Counter-signal: Project Mariner is dead. Google quietly shut down its 17-month browser-agent experiment May 4. The function got folded into Gemini Agent and Chrome auto-browse. The category that bet on screen-reading agents just lost its biggest sponsor. Command-line and API-shaped agents won the territory.

The two largest AI policy regimes in the world picked opposite directions in the same week. On Wednesday May 7, the EU Council and the European Parliament agreed provisionally to delay most of the AI Act's high-risk obligations by 18 months and slim the transparency grace period from six to three months. On Tuesday May 5, the US Center for AI Standards and Innovation finished sweeping every major US frontier lab into one pre-release safety-testing pipeline. Same week, opposite directions. Underneath that, the labor channel got noisier (Cloudflare 1,100, BILL up to 30 percent, Upwork 24 percent on Wednesday), the supply side added another 5 gigawatts (NVIDIA and IREN), and the industry pushed out the GPU networking spec it had been running in production. Project Mariner's quiet shutdown is the data point that ties the rest together: the agent shapes that won this week are not the screen-reading ones.

Brussels said wait. Washington said test. Cloudflare said cut. The same week.

1. The EU just delayed its AI rulebook by 18 months. The US tightened its own the same week.

Regulatory inflection. International on one side, federal on the other. On May 7, the Council of the EU and the European Parliament reached a provisional political agreement on Omnibus VII, the Commission's package of targeted amendments to the EU AI Act. The big move: the application of most high-risk AI obligations moves from August 2, 2026 to December 2, 2027, with safety components embedded in regulated products (Annex I) deferred further to August 2, 2028. The transparency grace period for general-purpose AI gets cut from six months to three, with the new deadline December 2, 2026. The Machinery Regulation gets a full carve-out from direct AI Act applicability where the two overlap. The package added one new prohibition: AI systems generating non-consensual intimate imagery and AI-generated child sexual abuse material.

Same week on the US side, the Center for AI Standards and Innovation (the renamed federal AI safety institute that lives at NIST inside the Department of Commerce) finished signing every major US frontier lab to pre-release national-security testing terms. Microsoft, Google, and xAI joined Anthropic and OpenAI under a single evaluation pipeline. The two regimes are now visibly diverging in cadence: the EU is buying compliance teams 18 more months while the US is making the federal red-team a routine pre-deployment step. For the lab procurement-and-vendor-management end of the org, that is two different timelines for the same vendors.

Council of the EU press release on the AI Act omnibus political agreement — consilium.europa.eu · May 7, 2026

Why this matters: If you are a solution architect or compliance owner running an EU rollout, the new high-risk deadlines (December 2, 2027 and August 2, 2028) reset your project plan. Standards, conformity assessments, and the national authorities you would have been registering with by August 2026 just slid out a year-plus. If you are deploying the same model in the US, you are dealing with the opposite shape: a federal pre-release evaluator (CAISI) that already gates the vendor before you see the model. Two timelines, same vendors, opposite trajectories. Watch the formal Parliament-and-Council adoption vote, currently expected later in May or June, before treating these dates as final.

https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/

2. Cloudflare cut 1,100 jobs and named a 600 percent surge in internal AI use. BILL went up to 30 percent. Upwork went 24 percent. Same day.

Continuing 5.07 #2 (PayPal 4,760 / Coinbase 700 cohort) with three new entrants in 24 hours. On May 7, Cloudflare CEO Matthew Prince announced the company is cutting about 1,100 jobs, roughly 20 percent of staff, the first mass layoff in Cloudflare's 16-year history. Stated cause: internal AI usage at the company has gone up more than 600 percent in three months, and the company is reorganizing for an agent-first operating model. Q1 revenue was $639.8 million, up 34 percent year-over-year, but the Q2 guide came in soft and the stock dropped roughly 24 percent after the print. The same morning, BILL told staff it is cutting up to 30 percent of headcount with a $30 to $60 million restructuring charge in the current quarter. Upwork CEO Hayden Brown announced cuts of about 145 employees, 24 percent of staff; Q1 revenue $195.5 million was up 1.4 percent year-over-year, the Q2 guide $190 million came in 6.9 percent below consensus, and the stock fell about 19 percent.

Three independent CEOs, three independent earnings cycles, three different AI usage stories, all converging on the same operating decision in the same 24 hours. The reads are not identical (Cloudflare's revenue is up sharply, Upwork's is barely growing, BILL is restructuring under pressure), but the answer they all reached for is the same: a smaller workforce that absorbs more of the work via internal AI use. Stanford Digital Economy Lab and others have started publishing pilot-to-production conversion figures on enterprise AI in the 11 to 14 percent range, well below CFO expectations. The companies announcing cuts are betting the curve bends faster from here.

CNBC coverage of Cloudflare's 1,100-job cut and Q1 2026 earnings — cnbc.com · May 7, 2026

Why this matters: If you are a support engineer, BA, or QA lead at a company that has not yet moved on agent rollouts, the public-company cohort is now large enough that "we are still evaluating" is going to read differently in your next planning meeting than it did a quarter ago. The Cloudflare line is the more uncomfortable one for IT pros to hear: internal AI use up 600 percent in three months is the kind of metric a leadership team will quote when they decide which roles to backfill and which not to. Treat it as the new internal benchmark to interrogate, not as a headcount strategy to copy. The pilot-to-production conversion gap (11 to 14 percent) is the gap your team is being asked to close.

https://www.cnbc.com/2026/05/07/cloudflare-net-q1-2026-stock-earnings-layoffs.html

3. NVIDIA and IREN added 5 more gigawatts to the AI build-out. NVIDIA paid for the right to invest up to $2.1 billion.

Structural deal. The "who can scale" question got another 5-gigawatt answer. On May 7, NVIDIA and Australian-listed data-center operator IREN announced a strategic partnership to deploy up to 5 gigawatts of NVIDIA DSX-aligned AI infrastructure across IREN's global pipeline over time. The flagship is a 2-gigawatt Sweetwater campus in Texas. As part of the deal IREN issued NVIDIA a 5-year warrant (a contract giving NVIDIA the right, but not the obligation, to buy shares at a fixed price) for up to 30 million IREN ordinary shares at a $70 strike, subject to regulatory approval. Total potential investment: up to $2.1 billion. IREN closed the May 7 session up roughly 19 percent in after-hours trading.

The shape of this is what to read closely. NVIDIA does not need to own IREN; it needs IREN's gigawatts to be running NVIDIA-flavored hardware. The warrant is the lever that ties the two destinies together without forcing a near-term acquisition. Across the last 60 days the same shape has shown up at SoftBank-Roze, OpenAI-Oracle, and now NVIDIA-IREN: a structural co-investment that locks in compute supply without consolidating the cap stack. The pattern itself is the story.

NVIDIA newsroom announcement of the IREN strategic partnership and Sweetwater campus — nvidianews.nvidia.com · May 7, 2026

Why this matters: For DBAs and infrastructure architects, the NVIDIA-IREN shape is the one to internalize. The compute under your future workloads is being booked years ahead, in 5-gigawatt chunks, by labs and chip vendors. When your platform team asks "where will our model inference run in 2027," the answer is increasingly a small set of pre-committed campuses. Treat the 5-gigawatt headlines as a forward calendar of what will and will not be available, not as financial trivia.

https://nvidianews.nvidia.com/news/nvidia-and-iren-announce-strategic-partnership-to-accelerate-deployment-of-up-to-5-gigawatts-of-ai-infrastructure

Tomorrow's signal lands here. Subscribe to PickBits Daily Signal if you want the operator brief, not the hype.

4. Six chip and lab vendors released the GPU networking spec they have been running in production.

AI breakthroughs. First jointly released open networking spec from this group. On May 7, OpenAI, NVIDIA, AMD, Broadcom, Intel, and Microsoft released Multipath Reliable Connection (MRC) through the Open Compute Project. MRC is an Ethernet protocol that spreads a single data transfer across hundreds of network paths and recovers from failures in microseconds. It extends the existing RoCE protocol (a way of running storage and compute traffic over Ethernet) with SRv6 source routing (a packet-steering technique). The headline number: the design supports clusters of about 131,000 GPUs in two switch tiers. The spec is already running in production at OpenAI's Oracle Cloud Abilene site and Microsoft's Fairwater supercomputers.

The interesting thing here is the joint release pattern. Six commercial competitors put a working, in-production spec on a neutral standards body the same day. That is the inverse of the pattern that has dominated GPU networking up to now (proprietary InfiniBand fabrics, vendor-locked optics). The training runs that the labs have been planning for 2027 needed a single shared substrate to be feasible. This is what the substrate looks like.

OpenAI announcement of the Multipath Reliable Connection networking spec via Open Compute Project — openai.com · May 7, 2026

Why this matters: For enterprise architects and platform engineers, MRC means the GPU networking stack you are going to depend on for 2027 training and inference jobs is converging on one open spec, with multi-vendor interop. That cuts a real risk off the roadmap (the "we picked the wrong fabric" risk). Track Open Compute Project for the reference implementations and any ratified profiles, because that is where your supplier short-list will be reading from.

https://openai.com/index/mrc-supercomputer-networking/

Δ The counter-signal - Project Mariner is dead. The browser-agent category just lost its biggest sponsor.

On May 4, Google quietly shut down Project Mariner, the 17-month experiment in screen-reading browser agents that used Gemini to navigate web pages the way a human would. The capability got folded into Gemini Agent and Chrome's auto-browse functions. The framing across coverage was "consolidation," but the practical answer is that a category Google itself had named and funded just got de-prioritized.

Track the contrast with the rest of the week. The agents that won territory in the same period are command-line and API-shaped: Claude Code, OpenAI Codex, the agentic operating modes inside Cursor and Replit, the agent framework underneath the MRC training stacks. Browser-as-canvas was the more demo-friendly approach; tool-call-as-plumbing turned out to be the one that scaled into product.

Android Authority report on Google's Project Mariner shutdown — androidauthority.com · May 5, 2026

Why this matters: If you have any multi-quarter pilots betting on a screen-reading browser agent as the integration shape, this is the week to revisit the architecture. The labs that have ridden the curve are the ones whose agents call APIs and run in containers, not the ones that drive a Chromium window. That is your supplier-shortlist read for the rest of 2026.

https://www.androidauthority.com/google-project-mariner-shutdown-3664323/

» What to watch this week

Formal EU Parliament-and-Council adoption vote on Omnibus VII. The May 7 deal is provisional. Watch for the formal vote schedule and the final wording on the December 2, 2027 deadline before treating the timeline as locked.
Q2 2026 earnings calls from the layoff cohort. Cloudflare, BILL, Upwork, plus PayPal and Coinbase from last week, will all need to put numbers on what the cuts bought. The Stanford 11-to-14-percent pilot-to-production figure is the bar.
The Anthropic CAISI carve-out question. Microsoft, Google, and xAI signed in. Anthropic has the older agreement. Watch for any update on whether the Anthropic terms get renegotiated to match the new entrants, and what that does to the Pentagon vendor list.
Open Compute Project ratification of MRC. The spec is released. Ratification, reference implementations, and vendor conformance tests are the next gate. Watch for AMD's and Broadcom's first-party drivers to see how fast it moves from spec to shipping silicon.
Whether Meta gets pulled into CAISI. Open-weight Llama and Muse Spark do not have a "release moment" the way a closed-API model does. Whether the federal evaluator finds a workaround is the next governance signal.

Tomorrow's signal lands here.