Workforce recomposition: what the 2026 AI layoffs actually mean for engineering

TL;DR. The 2026 layoffs at Cloudflare, Meta, Microsoft, and Coinbase are not cost cuts. They are workforce recomposition. The same dollars are flowing into AI infrastructure and into the engineers who rebuilt their workflows around it, while a category of support work quietly disappears. The 30 percent / 80 percent productivity gap inside companies is now wider than the gap between companies.

What did the 2026 AI layoffs actually announce?

In the same first quarter, Cloudflare posted $639.8 million in revenue, up 34 percent year over year, and cut roughly 1,100 jobs, about 20 percent of the company (Cloudflare Q1 2026 earnings call, 7 May 2026). Coinbase announced 700 layoffs, around 14 percent of headcount, on 5 May 2026. Microsoft cut about 6,000 in May 2025. Meta announced another 8,000 planned cuts in the same window.

These are not the layoffs of a downturn. Cloudflare's revenue grew faster than at any point in three years in the quarter it cut staff. Meta's 2026 capital expenditure guidance went up to $145 billion. The cash that left payroll did not leave the companies. It moved.

Are these cost cuts, or workforce recomposition?

The 2026 capital expenditure guidance across the four largest hyperscalers (Alphabet, Meta, Microsoft, Amazon) is approximately $650 billion. That figure is going almost entirely into AI infrastructure and into the people who can use that infrastructure to ship.

Read alongside the layoffs, the pattern is consistent. Microsoft cut about 6,000 jobs in May 2025 while accelerating job postings for AI engineers, Azure AI, and Copilot teams through the same quarter. Cloudflare's CEO Matthew Prince said publicly that the company will hire more people in 2027 than in 2026. The same dollar moved to a different role.

Calling this 'AI layoffs' is technically accurate and operationally misleading. The accurate name is recomposition. The dollars stay. The role profile changes.

Why is Cloudflare's CEO already planning to grow headcount in 2027?

On the 7 May 2026 earnings call, Matthew Prince was unusually direct about which roles were being cut and which would replace them. He said that over the past six months especially, the productivity gains from people directly talking to customers and directly creating code had been extraordinary, and that a lot of the support roles behind them were not going to be the roles that drove companies going forward.

Read carefully, that is a statement about which work compounds when AI is added to it. Customer-facing roles and code-producing roles compound. One engineer with the right workflow ships work that used to take a team. Support and coordination roles around them stop compounding, because the work they coordinated now runs without them.

The 2027 headcount target is higher than 2026 because productive roles produce more roles. They generate scope, projects, integrations, products. That demand has to be staffed.

What is the 30-versus-80 gap, and why does it matter more than the layoffs?

On Meta's Q4 2025 earnings call on 28 January 2026, CFO Susan Li reported that output per engineer had increased 30 percent since the start of 2025, with most of that growth coming from agentic coding tools in Q4 alone. For 'power users', the engineers who had actually rebuilt their workflow around AI, the increase was 80 percent year over year.

The 30 / 80 gap is the real story. Most engineers added AI on top of their existing workflow. They got 30 percent more done. A smaller group rebuilt the workflow around AI. They got 80 percent more done. That gap is sitting inside the same company, the same team, sometimes the same week of work.

The gap will widen. Tooling improves. Compounding improves. An engineer at 80 percent today is closer to 100 percent next quarter. An engineer at 30 percent gains less from the next model upgrade, because the bottleneck is the workflow, not the model.

What I keep coming back to: the gap between the 30-percent engineer and the 80-percent engineer at the same company is now larger than the gap between that 80-percent engineer and her next employer.

What does the engineer's role actually become?

The role moves from specialist who codes a slice, toward something closer to a generalist who orchestrates agents, owns an outcome, and understands the business context that outcome serves. Coinbase calls this the 'AI-native pod'. Brian Armstrong wrote on 5 May 2026 that some pods might be one person directing agents that cover engineering, design, and product responsibilities. The two-pizza team idea from twenty years ago, a team small enough to be fed with two pizzas, is being pushed to one slice.

Three things change in practice:

1. Outcome ownership replaces feature ownership. The engineer ships a working thing for a real customer, not a contribution to a feature backlog.
2. Agent orchestration becomes a skill. Writing prompts is not the skill. Designing the workflow the agents run inside, what they have access to, when they hand off, how their output is verified, is the skill.
3. Business context becomes table stakes. The engineer who knows why the work matters to the customer ships better work than the engineer who knows only what the ticket asks for.

This is more engineers, not fewer. It is engineers shaped differently from the engineers a 2023 job description was written for. The shape change is exactly what our agentic AI systems work for clinical workflows has been pulling teams toward, a quarter ahead of when it became a labour-market story.

Is there a road back to pre-AI engineering?

There is not, and the pattern from the last shift makes it clear why. When the cloud arrived in the early 2010s, companies that lifted-and-shifted virtual machines to AWS without using cloud-native services moved their data centres without moving their business. The companies that rebuilt around managed services, infrastructure-as-code, and serverless got the curve. The lifters-and-shifters spent another decade catching up.

The cycle this time is shorter. The companies investing $200,000 or $300,000 a year per team in AI tooling are doing the math one of our biggest clients did out loud last week: an enterprise license for a frontier AI coding tool costs a fraction of one engineer's salary, and used well by fifty engineers it produces more than two more engineers would. Buying the tool is not a cost decision. Firing without buying the tool is.

I expect the 2027 hiring waves to look similar to the 2014 cloud-native hiring waves. The engineers with the new shape will be hard to find, expensive, and worth it. The engineers without it will still find work, at a different tier of company and a different tier of compensation.

What does this mean for healthcare specifically?

The cost of writing code has dropped sharply. The work that actually puts AI to use inside a clinic has not.

The work that matters is building context around the workflow. The patient pathway. The handoffs between intake, clinical, and billing. The peak-hour rhythm. The data already in the EHR and the data that lives in someone's spreadsheet or notes app. Without that context, an AI capability has nothing to plug into. With it, the same capability becomes useful inside the clinic instead of staying conceptual on a slide.

Two things shift in practice once the context is in place.

Custom integrations replace SaaS add-ons. Most clinics still solve a workflow problem by adding another SaaS product to the pipeline. With cheaper code and clear workflow context, building a native integration directly into the clinic's existing systems is realistic for problems that did not justify a custom build before. The capability gets added without adding another vendor, another contract, another login. This is the pattern behind the private AI we deploy inside our clients' clinics: a native, governed capability rather than another SaaS layer on top.

The build fits the organisation, instead of the organisation fitting the build. Off-the-shelf clinical software is designed for the median clinic. With the workflow context already mapped, software can fit a specific clinic: its patient pathway, its existing handoffs, its peak-hour rhythm. That fit is what most clinical adoption quietly fails on. When the tool already matches the way the team works, adoption stops being a separate project. Our work with ROC Clinic on unified data integration sits in that pattern.

The recomposition pattern reshaping Cloudflare and Coinbase applies to healthcare too. It arrives a quarter later, the constraints are different (UK GDPR, NHS DSPT, on-premise residency), and the pattern is the same. The work is the context.

How do you end up on the right side of this curve?

I'd be making this observation about other companies if I weren't watching it happen inside our own. Three moves, in the order we made them at GreenM. Each one taught us something I'd pass on.

1. Rebuild your weekly synthesis work first. Every weekday morning a routine scans my Outlook (sent and inbox), Slack across the channels I care about, and the meeting notes our agent has assembled from the past 24 hours. It cross-references everything against my Linear board, updates statuses where reality has moved, enriches descriptions with fresh context from the actual conversations, and creates new tasks for items I forgot to capture. Then I run a voice review: 15 minutes of talking through the board, pushing decisions back to Linear in real time. The Monday catch-up that used to take 90 minutes of clicking now takes 15 minutes of voice, and the board is more honest because it is grounded in what actually happened in the threads, not what I remembered.

The transferable point: if you have one workflow worth rebuilding first, this is the shape. The multi-tool synthesis work you do every week that lives in too many places to track is the highest-yield first target, on any team I have watched try this.

2. Buy the tooling instead of cutting the headcount. The math one of our biggest clients did out loud last week: an enterprise license for a frontier coding tool costs a fraction of one engineer's salary, and used well by fifty engineers it produces more than two more engineers would. They moved their teams onto enterprise licenses with a $150-per-engineer-per-month overage budget. By day three, several engineers had blown through the cap. The VP didn't tighten the cap. He read it as a signal that the engineers were using the tool the way the tool was meant to be used.

The transferable point: a team spending $200,000 a year on AI tooling and keeping its engineers will out-ship a team that cut two engineers to save the cash. The compounding is in what those engineers ship with the tooling. Our AI Launchpad is the eight-week version of that bet for healthcare teams getting started.

3. Move the hiring bar, and rewrite the role. Shopify's Tobi Lütke wrote on 7 April 2025 that managers asking for more headcount have to prove the work cannot be done with AI before they get approval. Eight months later, that standard is spreading. Inside GreenM we have rewritten the competency framework for five engineering roles. Every role now lists business understanding, agent orchestration, and direct work with the client as evaluated capabilities, alongside the technical depth that has always been there. The point is not to filter people out. It is to be honest with everyone, including ourselves, about what the role actually requires in 2026.

The transferable point: the role description on most teams is still the 2023 one. Update it before the market does it for you.

Frequently asked questions

Are large tech companies hiring fewer engineers in 2026?

No. They are hiring differently-shaped engineers. Cloudflare's CEO publicly guided that 2027 headcount will be higher than 2026. Microsoft's net engineering headcount was roughly flat to slightly up year over year in Q1 2026 even after the May 2025 layoffs. The shape of the role is changing, not the count.

What's the difference between using AI in a workflow and rebuilding the workflow around AI?

Using AI in a workflow means adding a prompt to an existing step. Rebuilding the workflow around AI means changing which steps exist. The Meta 30 / 80 gap maps directly to this distinction. The first move gets you 30 percent. The second gets you 80 percent.

How fast is the 30 / 80 gap widening between engineers?

Faster than model release cycles. Power users compound on each model release. Non-power users do not, because their workflow does not let the new capability through. From what I have seen with our clients across 2025 and into 2026, the gap roughly doubles every two quarters.

Does this apply to small companies and consultancies, or only to Big Tech?

It applies more sharply to small companies, because there are fewer roles to hide work inside. Big Tech can recompose over multiple quarters. A 30-person consultancy recomposes in weeks, by the next contract.

What's the one workflow most engineering teams should rebuild first?

Whatever you currently do every Monday morning by reading across three or four tools to figure out where things stand. That synthesis work is the highest-yield first target. It is repeatable, it is mechanical, and it is the source of the most procrastination on any team I have observed.

What this leaves us with

Cloudflare did not cut 1,100 people because AI got cheaper. It cut 1,100 because the shape of the company changed and the support roles around the productive ones stopped compounding. Meta is not measuring its 80-percent power users because they are heroes. It is measuring them because the gap between them and the median engineer is the gap between this year's company and next year's.

The 2026 layoffs are the headline. The recomposition is the story. The engineers who get to the other side of it are not the engineers who were laid off. They are not the engineers who were not laid off either. They are the ones who rebuilt their workflows around the tools while the rest of the industry was still arguing about whether the tools mattered.

If you build one workflow this quarter that you actually run every week, that is the bet that pays back. The compounding starts now.

Sources

1. Cloudflare, Inc. (NET) Q1 2026 Earnings Call Transcript, 7 May 2026. The Motley Fool.
2. Meta Platforms (META) Q4 2025 Earnings Call Transcript, 28 January 2026. The Motley Fool.
3. Brian Armstrong / Coinbase to lay off 700 workers as CEO restructures for AI efficiency, 5 May 2026. Fortune.
4. Tobi Lütke on X: 'Reflexive AI usage is now a baseline expectation at Shopify', 7 April 2025.
5. Cloudflare CEO Matthew Prince on layoffs and AI automation, 21 May 2026. Fortune.
6. Big Four AI capital expenditure guidance, Q1 2026 disclosures. Fortune.