Welcome to Level.UP, brought to you by UP.Labs.

This week, we dive into two stories.

The first: Anthropic and OpenAI are both investing in enterprise AI services firms, capitalized by private equity. The second: a closer look at how China's approach to AI — diffusion over AGI, applications over models, factories over chatbots — is producing a deep operational moat that operators should be paying attention to.

Plus: tools we're tracking and hot takes on recent physical AI news.

Think someone else needs this? Forward it to a friend or colleague navigating the same terrain.

Anthropic and OpenAI both announced enterprise AI services ventures within hours of each other earlier this May.

The structures are nearly identical: the AI provider supplies the models and engineers, private equity supplies the capital and a captive client base, and the joint venture deploys both into mid-sized companies that can't build AI capability on their own.

Anthropic's firm, backed by roughly $1.5 billion from Blackstone, Hellman & Friedman, and Goldman Sachs, embeds Anthropic engineers alongside the new firm's own team. General Atlantic, Apollo, Leonard Green, GIC, and Sequoia round out the cap table.

OpenAI's Deployment Company raised $4 billion from a 19-firm syndicate at a $10 billion pre-money valuation, led by TPG with Brookfield, Bain Capital, and Advent as co-leads. Notably, OpenAI's investor list also includes Bain & Company, Capgemini, and McKinsey — three of the consultancies whose business model the venture is built to undercut.

Services are becoming the new software.

Sequoia partner Julien Bek's thesis, which has circulated widely since he posted it earlier this year, gives the clearest framing of what's happening. For every dollar enterprises spend on software, they spend six on services. The next great company, in his view, won't sell software at all but outcomes — legal work, financial analysis, claims processing — delivered by AI and billed like consulting.

Blackstone's Jon Gray named the underlying constraint plainly. The goal of the new firm is to break down "one of the most significant bottlenecks to enterprise AI adoption" — the scarcity of engineers who can actually implement frontier systems at speed.

But the fundamental trade hasn't changed. Engineers come in, study your operation, and bolt AI onto the systems you already run. You get a smarter version of the operation you already had. For many companies, that's the right move. The systems work, the customers know them, and the goal is to make them run better. "AI-fying" what you've already built is a real path to value.

There's a gap this path leaves open, though, and it's the one we work in at UP.Labs. 

Some problems are bigger than the operation they currently live inside. They don't get solved by making the existing system smarter. They get solved by building a new company around what's actually possible now. That's the venture-build path: instead of retrofitting AI onto a legacy operation, you build an AI-native company around the problem the legacy operation was solving, designed from the start around what these systems can actually do.

It requires real partnership between the operator and the builder. But it produces something the retrofitting path can't: a company whose operations are designed for the technology running them, with the upside of owning equity in the result rather than paying a services bill.

While most US coverage of AI fixates on chatbots, model benchmarks, and the race to AGI, China has been playing a different game for the past 15 years.

The differences came into focus on a recent podcast that examined how thoroughly AI is now embedded in everyday Chinese life. Cars, restaurants, video platforms, retail, and consumer services all run on AI systems built on top of the country's enormous installed base of automation.

That density of deployment generates something Western AI companies are starting to run short of: training data. 

While US labs worry about exhausting the open web, Chinese systems are continuously trained on data from robotaxis, smart factories, surveillance networks, and consumer apps used by hundreds of millions of people. China has a significant structural edge in physical AI, thanks to its overlapping hardware ecosystems and its ability to manufacture at scale.

Industrial AI penetration in Chinese enterprises reportedly jumped from 9.6% to 47.5% in a single year, according to the most recent CNNIC report.

The most important advantage in AI isn't the model. It's the data the model learns from.

That's where China's position becomes hard to compete with. Every robotaxi ride, every Foxconn production line, every transaction inside a state-run platform produces operational training data. While Western models are scraping the open web and worrying about running out of data, Chinese models are being fed by the physical economy.

State control compounds it. Chinese programs like Skynet aggregate traffic, policing, and citizen-movement data at a scale no Western company could legally assemble. And when a strategically important company drifts outside Beijing's reach, the state pulls it back. The Meta-Manus merger block is the latest example.

For US operators in physical industries, this is the part worth paying attention to. China's 15-year head start as the factory of the world is now compounding through AI, as every factory floor and every shipped product trains the next model. That's a different kind of moat than anything being built in a Bay Area data center.

That’s why these are the questions we’re asking companies: where is your operational data going, who is structuring it, and is anyone treating it as the strategic asset it's becoming? Most legacy firms are excellent at their core business and have never realized that their decades of sensor logs, yield curves, and process telemetry are an invaluable asset in and of themselves.

In practice, we’ve seen at least three ways you can leverage this data: mine it as internal IP to solve problems that exist up and down your value chain; pair it with a growth-stage startup's hardware or models in a joint venture; or license anonymized data to other firms training models and building digital twins.

NVIDIA is already running this play. Jensen Huang, NVIDIA’s CEO, has been explicit that the company is going industry by industry, pairing its stack with companies’ operational data because that's where the real edge lives. The real opportunity for US industrials isn't catching up on models. It's recognizing the goldmine that’s already in your factory or data warehouse.

In the last edition, we shared the first few internal apps Jesse Silverman built to streamline our recruiting process. This week, we have two more to share, along with video walkthroughs of how each one works.

Neither of these started as an "AI strategy." They started as a list of things one person was doing manually, 10 to 15 times a week, that a model could do better with the right context wired in.

WIRobotics Raises $68M For Humanoids. The South Korean robotics company recently closed a Series B to scale its humanoid platform, ALLEX. But the asset isn't the robot — it's three years of gait data from WIM, its wearable walking-assist device. Since 2023, it’s sold 3,000+ units across five markets and has grown revenue 5x. The signal: in physical AI, the moat is rarely the form factor. It's the operational data that drives it. → Read more

Open-Source Hijack Hits OpenAI. OpenAI disclosed last week that two employee devices were compromised after attackers hijacked TanStack, a widely used open-source code library that thousands of companies — including OpenAI — pull into their software. The campaign has also reached libraries from Mistral AI, UiPath, and OpenSearch. OpenAI is rotating its code-signing certificates and requiring all macOS users to update by June 12. The lesson: most AI tools run on open-source components that the company didn't write and doesn't fully audit. That dependency layer is the highest-leverage attack surface in enterprise AI today. → Read more

Faraday Future Bets Its Comeback On Robots. The Gardena-based EV maker raised $25M in convertible notes to fund a pivot toward embodied AI — humanoid, bionic, and automotive robots — with a stated goal of restoring its 2021 IPO valuation and reaching positive operating cash flow by Q4 2027. The read: physical AI is now attracting both pure-play builders like Physical Intelligence and Anvil Robotics, as well as legacy companies repositioning to access the category's capital. → Read more