286 People. 800,000 Applications. $2.57 Million Revenue Per Employee.

“A 286-person company received 800,000 job applications in 2025, hired 286 of them, had AI write 90% of its pull requests by Q1 2026, and generated $2.57 million in revenue per employee — more than double the prior year. Bending Spoons is not a thought experiment. It is the operating model that every AI-first business will eventually converge on. The question is whether your business is moving toward it or waiting to see what happens.”

The Bending Spoons Profile: Every Number That Matters

The Wall Street Journal published a profile of Bending Spoons this week that AI Weekly’s coverage flagged as the most widely shared enterprise AI story of the week. Here are the specific numbers from the profile:

• 286 hires from 800,000 applications in 2025 — a 0.036% acceptance rate, narrower than Harvard
• 90% of pull requests written by AI by Q1 2026
• Revenue per employee more than doubled year-over-year to $2.57 million
• Total headcount: 286 people managing a portfolio of software products used by millions of users globally
• The hiring ratio is driven by philosophy, not scarcity: Bending Spoons deliberately keeps its team small and compensates at rates that make each hire economically equivalent to multiple conventional employees

Let that last number sit: $2.57 million in revenue per employee. The average for S&P 500 technology companies is approximately $500,000 to $700,000 per employee. Bending Spoons is operating at 3.7 to 5x the revenue efficiency of a typical technology company — not at 10 to 20% better, but at a fundamentally different order of magnitude.

What Bending Spoons Actually Does

Bending Spoons is an Italian software company (headquartered in Milan) that acquires consumer software products and applies operational excellence to scale them. It acquired Evernote, Splice, StreamYard, WeTransfer, and several other consumer software products. Its business model is not building software from scratch but buying existing products and improving their unit economics through better product development and lower operational overhead.

The AI-first operating model that produced 90% AI-written pull requests and $2.57 million revenue per employee is not a research project. It is the operating model applied to a real portfolio of products with real revenue. That is the detail that makes the Bending Spoons profile more credible than most AI productivity claims: these numbers come from a company with external auditors, public financials on its debt instruments, and millions of paying customers whose retention data validates product quality.

What 90% AI-Written Pull Requests Actually Means in Practice

The 90% figure is the one circulating most widely, and it deserves careful interpretation. In the AI developer community, “AI-written code” covers a spectrum from an autocomplete suggestion that fills in a function body to a Claude Code autonomous agent that plans, implements, tests, and creates a complete pull request with zero human keystrokes in the implementation phase.

The Bending Spoons figure, per the WSJ profile context, refers to code where AI handled the primary implementation: the developer defined the specification, reviewed and approved the output, and shipped it — but the code was generated by AI rather than typed by the developer. The human role shifted from implementation to specification and review. That shift is the one that allows 286 people to maintain a software portfolio at the quality and scale that would previously have required thousands of engineers.

This is the automation ratio concept applied to software engineering: the percentage of pull requests that ship with AI-generated implementation and human review, rather than human-generated implementation and human review, is Bending Spoons’ engineering automation ratio. At 90%, they have achieved a level of trust in AI-generated code output that only 10% of PRs require human implementation. That is not an accident. It is the product of deliberate workflow design, review discipline, and quality measurement that took years to build.

The 800,000-to-286 Hiring Filter

The 0.036% acceptance rate is not primarily a quality statement about rejected applicants. It is a statement about the leverage Bending Spoons requires from each person it hires. A team where AI handles 90% of code implementation needs human contributors who can do the 10% that AI cannot — architectural judgment, product taste, specification precision, and the quality evaluation that determines whether the 90% AI output is good enough to ship.

The most valuable skill in a 90%-AI-code environment is not writing code. It is the ability to specify what correct code looks like and to recognise when AI-generated code is subtly wrong in ways that will cause problems later. That skill is rarer and more valuable than implementation ability, which is why the hiring ratio is so tight and the compensation is so high.

This validates the Satya Nadella argument about human capital: human capital does not become less valuable as AI capability grows. The nature of the valuable human capability shifts. Bending Spoons is the most concrete existing demonstration of what that shift looks like in a company that has already crossed the 90% threshold.

The Bloomberg Dot-Com Parallel: Why This Rally Might Split

Bloomberg published an analysis this week drawing a parallel between the current AI market and the dot-com era’s market split. Per the Bloomberg analysis via AI Weekly, the AI rally is more likely to cleave the trade into infrastructure versus software than to kill it, particularly as Iran threat concerns create market uncertainty. The infrastructure companies (Nvidia, SK Hynix, data centre REITs) are viewed as direct plays on compute demand that is visible and contracted. The software companies (OpenAI, Anthropic, enterprise AI SaaS) face the same question the dot-com era’s software valuations faced: how quickly does revenue grow into valuation multiples that currently rely on future expectations?

The Infrastructure vs Software Split in Practice

The Bloomberg split thesis is already visible in the market data. SK Hynix at $1 trillion on day one of trading, as covered in today’s latest article, is infrastructure. Etched launching at a $5 billion valuation with $1 billion in signed contracts for specialised AI inference chips is infrastructure. The Gartner $206.5 billion AI agent software spending forecast and the AI software companies approaching public markets are software. Infrastructure has visible, contracted revenue. Software has rapid growth curves that are real but valued on expectations.

For businesses making AI investment decisions: the infrastructure-software split is relevant to your vendor selection risk as well as to investment portfolios. Infrastructure-layer vendors (compute providers, memory manufacturers, chip companies) have pricing power rooted in physical supply constraints. Software-layer vendors have pricing power rooted in switching costs and capability advantages that can compress as competition intensifies. Both are real. They are different risk profiles.

Sam Altman vs Elon Musk: The $2 Trillion Valuation Fight

Sam Altman publicly accused Elon Musk of “selling public market investors on short-term space data centers” amid the $2 trillion SpaceX valuation push, per AI Weekly’s coverage. The specific allegation: Musk is marketing SpaceX’s data centre capacity as an AI infrastructure play to sustain the $2 trillion valuation, when the actual AI workloads that data centre capacity will run are not yet contracted at the scale the valuation implies.

Altman’s accusation is self-interested: OpenAI competes with xAI for AI developer mindshare, and Musk’s Grok models and SpaceX Colossus compute are direct competitive assets. That does not make the underlying argument wrong. The Bloomberg dot-com split thesis applies directly: SpaceX’s $2 trillion valuation requires a view of AI data centre demand that may or may not materialise at the contracted scale the valuation implies.

For the automation builder community: this Altman-Musk fight is primarily a financial markets story. Its practical implication is the same one the Bloomberg analysis surfaced: be cautious about computing infrastructure lock-in with any single provider whose valuation is substantially built on AI demand projections that are not yet in long-term contracts. The providers with contracted, visible demand are structurally safer dependencies than those whose valuation is built on future demand expectations.

The Bending Spoons Model: What It Means for Every Business

The Bending Spoons profile is the most concrete available demonstration of what

the ‘AI business advice is wrong’ thought leadership piece argued: the businesses building durable competitive advantage with AI are the ones that redesigned their operations around AI capability rather than adding AI tools to existing operations.

Bending Spoons did not add AI to its existing software development workflow. It redesigned the software development workflow around AI capability: hired for specification and judgment skills rather than implementation skills, achieved 90% AI implementation, and freed the people it did hire to focus entirely on the work that AI cannot do.

The Specific Lesson for Service Businesses

Bending Spoons is a software product company. The operating model translates differently for service businesses, but the underlying principle is identical: identify which tasks in your service delivery AI can handle at acceptable quality, redesign the workflow so AI handles those tasks systematically, and retain human talent for the judgment and specification work that determines whether AI output is good enough.

At Hexona Systems, we are not at 90% AI-generated deliverables. We are building toward it for specific workflow categories. The client-facing strategy work, the diagnostic conversations, the outcome definition — those require human judgment that I do not see AI replacing in the near term. The implementation, the documentation, the reporting, the quality checks — those are progressively automatable, and we are automating them progressively. The measure of progress is the same as Bending Spoons: revenue per person, and the proportion of deliverables that reach the client without a human implementing them from scratch.

The Labour Market Implications

The Bending Spoons model, applied across the economy, is what the Canaries Dashboard finding is measuring: workers aged 22 to 25 in AI-exposed occupations are experiencing 3.8% annual employment decline, while senior workers with specification and judgment skills are relatively protected. The Bending Spoons hiring model is precisely the one that creates this pattern: 286 senior hires from 800,000 applicants, with AI handling the implementation that junior hires would previously have provided.

The June jobs report showing only 57,000 jobs added against a 185,000 consensus, with 88,000 AI-attributed job cuts in 2026, is partially the aggregate of thousands of companies moving toward Bending Spoons-style hiring ratios in their own ways. No single company is Bending Spoons, but a broad directional shift toward ‘fewer people, higher specification skills, more AI implementation’ produces the macro numbers we are seeing.

Meta Compute: Zuckerberg Monetises the AI Overbuild

Meta unveiled ‘Meta Compute’ this week — a plan to rent out its AI data centre capacity to external customers, per the Motley Fool via AI Weekly. Zuckerberg is monetising the infrastructure overbuild that has drawn investor criticism: Meta has spent enormous capital on AI infrastructure, and now that infrastructure can generate cloud computing revenue alongside its internal AI applications.

The Meta Compute announcement has two implications for businesses building AI automation. First, AWS, Google Cloud, and Azure will have a fourth major competitor for enterprise AI compute within 12 to 18 months, which is good for pricing competition in cloud AI infrastructure. Second, Meta’s compute infrastructure runs on Llama models at its core, which means Meta Compute customers will have incentive to use Llama-family open-weight models that are optimised for that infrastructure, potentially at pricing advantages over frontier model APIs.

For businesses currently building on AWS, Google Cloud, or Azure: the Meta Compute announcement is a reason to maintain portable architecture and avoid deep infrastructure lock-in with any single provider before the competitive landscape fully settles. The AI agent platform war analysis’s core argument — flexible contracts, portable architecture, no single-vendor lock-in — applies to the infrastructure layer now as much as to the model layer.

The Week’s Pattern: Concentration and Divergence

The week of July 6 to 13, 2026 has a coherent pattern underneath its individual stories: the AI market is simultaneously concentrating at the top and diverging in the middle.

Concentration: 88% of AI startup funding goes to US companies, per Crunchbase H1 data, and most of that to a small set of frontier labs. SK Hynix at $1 trillion, Nvidia’s continued dominance, and the Apple-OpenAI lawsuit are all expressions of winner-take-most dynamics at the frontier.

Divergence: Bending Spoons at $2.57 million revenue per employee, 30-46% of enterprise tokens flowing to Chinese open-weight models, and the Etched inference chip at $5 billion without being a frontier lab are all expressions of a more distributed market in the middle tier. The businesses that understand this divergence — and position accordingly, using frontier capability for the specific tasks that justify it while capturing the efficiency of the middle tier for everything else — are the ones building durable advantage in this environment.

The Bottom Line

The Bending Spoons profile is the most practically important story of the week because it is the first rigorously sourced, publicly documented case study of what a fully AI-first operating model produces at company scale: $2.57 million revenue per employee, 90% AI-generated pull requests, 286 people managing a multi-product software portfolio used by millions. It is not a thought experiment. It is what the endpoint of AI-first operations looks like for a software business.

Most businesses will not reach this endpoint in the near term. Most will not try to. But understanding that this endpoint exists, and that a 286-person company is already operating there, should change how every business owner thinks about the direction they are building toward. The question is not ‘Should we use more AI?’ The question is ‘How many of the tasks in our business could AI handle at acceptable quality if we designed the workflow around that goal?’

Start counting. The five-question classification framework and the automation ratio metric are the tools. The Bending Spoons profile is the evidence that the direction is correct.

Frequently Asked Questions

What is Bending Spoons and why is the WSJ profile significant?

Bending Spoons is an Italian software company that acquires consumer software products and applies AI-first operations to scale them. The WSJ profile revealed it made just 286 hires from 800,000 applications in 2025, had AI write 90% of its pull requests by Q1 2026, and generated $2.57 million in revenue per employee — more than double the prior year. The significance: these are audited, external revenue numbers from a real company with paying customers, not a projection or a case study. It is the first publicly documented, rigorously sourced example of what a fully AI-first software operating model produces at company scale.

What does ‘90% of pull requests written by AI’ actually mean?

At Bending Spoons, AI handled the primary implementation of 90% of pull requests: the developer defined the specification, reviewed and approved the output, and shipped it, but the code was generated by AI. The human role shifted from writing code to specifying what correct code looks like and evaluating whether AI-generated code meets that specification. This shift is what allows 286 people to maintain a multi-product software portfolio at a quality and scale that would previously have required significantly more engineers.

What is Bloomberg’s dot-com market split thesis for AI?

Bloomberg’s analysis argues that market disruptions are more likely to split the AI trade into infrastructure versus software than to end the AI rally. Infrastructure companies (chipmakers, compute providers, data centres) have visible, contracted revenue rooted in physical supply constraints. Software companies have rapid growth that is real but valued on future expectations. This creates different risk profiles: infrastructure is valued on contracted demand, software on projected demand. Both are legitimate investments, but the divergence between them is likely to widen as market conditions tighten.

What is Meta Compute and how does it affect the cloud AI market?

Meta Compute is Meta’s plan to rent out its AI data centre capacity to external customers, monetising the infrastructure overbuild that has drawn investor scrutiny. It creates a fourth major competitor for enterprise AI compute alongside AWS, Google Cloud, and Azure. This is good for pricing competition in cloud AI infrastructure. Meta Compute customers will likely benefit from pricing advantages on Llama-family open-weight models optimised for Meta’s infrastructure. The timeline for Meta Compute’s commercial availability has not been officially confirmed.

Can a service business achieve Bending Spoons-level AI efficiency?

Not at the same ratios, and not on the same timeline. Bending Spoons’ 90% AI-written PRs is specific to software implementation, which has well-defined correctness criteria and automated testing for validation. Service businesses have more variable output quality criteria and less automated validation of AI-generated work. That said, the directional principle is identical: identify which specific tasks AI can handle at acceptable quality, redesign the workflow so AI handles them systematically, and measure progress using the automation ratio. The endpoint for a service business is different from Bending Spoons, but the measurement framework and the direction are the same.

Related Reading From This Series

The ‘AI Business’ Advice Is Wrong — why Bending Spoons’ model vindicates process ownership over tool delivery

The Automation Ratio — the metric that measures your progress toward Bending Spoons-level efficiency

You Don’t Need an Agent. You Need a Rule. — the classification framework for deciding which of your tasks should be AI-generated

Satya Nadella’s Learning Loop Warning — why Bending Spoons’ hiring ratio is the enterprise expression of human capital becoming more, not less, valuable

30-46% of US Enterprise Tokens Flow to Chinese Models — the middle-tier market that sits between Bending Spoons’ AI-first model and frontier-only approaches

80% of Executives Report No Measurable AI ROI — why the 80% are not Bending Spoons and what the difference is

Apple Sues OpenAI, SK Hynix Hits $1 Trillion — the infrastructure concentration story running alongside Bending Spoons’ software efficiency model

About the Author: Hamza Baig is the founder of Hexona Systems, an AI automation agency serving clients across six continents, and creator of the AI Automation Institute, where over 40,000 entrepreneurs have learned to build and scale automation businesses. He has been featured in GHL Top 50, Yahoo Finance, and Brainz Magazine. Follow him at @hamza_automates | Read more articles | Work with Hamza