The Day Everything Changed

May 4, 2026. If you blinked, you probably missed it. But that Monday in Silicon Valley, something genuinely interesting happened. Two companies that spend most of their waking hours trying to destroy each other — OpenAI and Anthropic — announced almost identical moves on the exact same day.

They both decided to marry into Wall Street wealth. OpenAI pulled together 19 investment firms, including TPG, Bain Capital, Brookfield, and SoftBank, raised over $4 billion, and launched a new entity called The Deployment Company — DeployCo for short, valued at $100 billion. Anthropic did the same, partnering with Blackstone, Goldman Sachs, and Hellman & Friedman, raising around $1.5 billion for a similar venture.

Same day. Same strategy. Same target: helping companies actually get AI to work inside their real-world operations.

This isn’t just another funding round. This is OpenAI admitting that selling API tokens and enterprise seats isn’t enough anymore. The game has changed.

From Software Vendor to Full-Service Consultant

Here’s what DeployCo actually does. It’s not complicated, but it is expensive. OpenAI is sending forward-deployed engineers — FDEs, in industry speak — directly into client offices. These aren’t junior support staff. These are AI experts who sit in your conference rooms, flip through your process documents, interview your frontline employees, extract data from your legacy ERP systems, and rebuild your workflows from the ground up using OpenAI’s models.

Think of it as Palantir’s playbook, but for AI. Palantir grew into a defense and finance powerhouse by embedding engineers directly inside client organizations, learning their business language, and writing code on the fly. Now OpenAI is doing the same thing, but with large language models as the core weapon.

To turbocharge this effort, OpenAI also acquired Tomoro, a London-based AI consulting and engineering firm, bringing about 150 deployment specialists directly into the fold. These engineers will be the boots on the ground, working with clients like BBVA, the Spanish bank that just signed on as a founding partner, along with Goldman Sachs and a consortium of other financial and consulting heavyweights.

Goldman’s Marc Nachmann put it bluntly: “What’s missing right now isn’t the technology itself. It’s knowing how to apply these tools to real business scenarios.”

He’s right. For the past three years, the AI industry has been obsessed with model benchmarks — who has the biggest parameter count, the lowest hallucination rate, the highest MMLU score. Meanwhile, business leaders have been looking at their employees using ChatGPT to write emails and wondering: is this actually saving me money?

The answer is yes, but it’s hard. Really hard.

The Unspoken Problem No One Wanted to Admit

You can’t just plug AI into a supply chain that’s been running on spreadsheets for fifteen years. You can’t hand a large language model to a customer service team and expect it to understand your product logic and brand voice without extensive customization. These aren’t API calls. These are organizational transformation projects.

And the people who can do this work — engineers who understand both frontier AI models and messy real-world business processes — are the scarcest talent on the planet.

OpenAI and Anthropic realized something that Salesforce figured out twenty years ago and Palantir figured out fifteen years ago. Standardized products never cover the last mile of enterprise adoption. If you want big customers to actually use your software, you have to send people to hold their hands.

But here’s where it gets really interesting. Why are both companies partnering with private equity giants instead of traditional consultancies like Accenture or Deloitte?

Because PE firms hold the keys to the kingdom. Blackstone, TPG, and their peers own thousands of portfolio companies across healthcare, manufacturing, finance, retail, and real estate. These mid-sized enterprises have data, they have use cases, they have budgets — but they don’t have AI engineering teams. They’re exactly the customers that AI vendors want most and struggle to reach.

By partnering directly with the PE firms that own these companies, OpenAI and Anthropic get instant access to a massive, high-quality, captive customer network. The PE firms get to offer their entire portfolio a group discount on AI transformation. If a fund can increase efficiency at just one portfolio company by 10 percent, the valuation lift could dwarf the hundreds of millions they’re investing in these ventures.

It’s a classic win-win, executed at an unprecedented scale.

Why Now? The Pressure Behind the Pivot

This isn’t happening in a vacuum. OpenAI is feeling real heat for the first time since ChatGPT launched.

Let me give you the numbers that should scare anyone at OpenAI. According to new data from Enterprise Technology Research, OpenAI’s enterprise market share peaked at 62 percent in September 2025. By March 2026, it had dropped to 56 percent. That’s still number one, but the lead over Anthropic has collapsed from 41 percentage points a year ago to just 8 points today.

Meanwhile, Anthropic’s enterprise adoption more than doubled from 21 percent to 48 percent in twelve months. First-time AI buyers are now choosing Anthropic at three times the rate of OpenAI. Claude’s average revenue per active user is $16.20, miles ahead of OpenAI’s $2.20 and Google’s $1.10. And here’s the kicker — in Q1 2026, Anthropic actually overtook OpenAI in total LLM revenue, capturing 31.4 percent of the market compared to OpenAI’s 29 percent.

Let that sink in. Anthropic, with far fewer users than OpenAI, is making more money. They cracked the code on enterprise monetization while OpenAI was busy chasing consumer scale.

Sam Altman saw this coming months ago. He reportedly issued a “code red” memo late last year, realizing that OpenAI had spread itself too thin — chasing everything from video generation models like Sora to robotics experiments — while Anthropic stayed laser-focused on business customers and coding tools. The strategy backfired.

Now Altman is course-correcting hard. He plans to double OpenAI’s workforce to 8,000 employees by the end of the year, with new hires across product development, engineering, research, and sales. OpenAI is also winding down its so-called “side quests” to concentrate on core enterprise products. Fidji Simo, OpenAI’s applications CEO, told employees: “We must not miss this moment because we are distracted by side quests.”

The Revenue Problem That No One Talks About

OpenAI has 900 million weekly active users. That’s an astonishing number. But 90 percent of them don’t pay a cent. The company’s average revenue per user is just $2.20 per month. For context, Netflix makes about $15 per user. Spotify makes about $10.

The consumer subscription model works for scale, but it doesn’t generate the kind of margins needed to sustain the AI infrastructure arms race. According to financial forecasts circulating on Wall Street, OpenAI’s annual compute spending could hit $121 billion by 2028, with annual net losses exceeding $85 billion.

That’s not a typo. Eighty-five billion dollars in losses. In one year.

This is why OpenAI is simultaneously pushing into advertising — launching a self-serve ad platform with CPC billing, targeting $2.5 billion in ad revenue in 2026 — and exploring “value-sharing” models where they take a cut of customer revenues generated by AI. Sarah Friar, OpenAI’s CFO, floated this idea in Davos: if a pharmaceutical company uses OpenAI models to discover a drug that generates billions in sales, OpenAI wants a piece of that action.

It’s the difference between selling shovels and owning a stake in the gold mine.

What DeployCo Means for the Industry

The formation of DeployCo represents something bigger than just another OpenAI product launch. It’s an admission that the pure software model has limits in enterprise AI.

Selling API access has gross margins of 80 percent or more. Selling deployed services has gross margins of 30 to 50 percent. From a pure financial perspective, this looks like a step backward. But the market values software-plus-service hybrids differently — not just on current margins, but on lock-in and lifetime customer value.

Once OpenAI’s FDEs rebuild your core business processes around their models, you’re not switching to Anthropic next quarter. The switching costs are enormous. You’ve rebuilt supply chains, retrained staff, and rewired integrations. That customer is sticky for years.

This is the play. Accept lower margins upfront to capture the most valuable, deepest integrations, then collect tolls for the next decade.

It also puts OpenAI in direct competition with a whole new set of players. Indian IT services giants like Tata Consultancy Services, Infosys, and HCL Technologies saw their stocks tumble 5 to 6 percent on the DeployCo announcement, hitting new 52-week lows. These companies built billion-dollar businesses on labor arbitrage — deploying armies of engineers to client sites for application development and maintenance. Now OpenAI is coming for that same model, but with AI-native engineers who can do in weeks what used to take months.

The Multi-Model Reality

One more thing worth watching. Enterprises are no longer betting on a single AI vendor. A16z’s latest CIO survey found that 81 percent of large companies now use three or more model families, up from 68 percent just a year ago. The model capability gap that dominated the conversation in 2024 has largely closed for enterprise use cases. Now the battle is commercial, contractual, and operational.

This means OpenAI can’t just rely on GPT being the best model. They have to win on service, on integrations, on trust, on the entire package.

DeployCo is their answer to that new reality. It’s not about having the smartest model anymore. It’s about being the vendor that can actually get the damn thing working inside a real company, with real data, real compliance requirements, and real people who just want their workflows to stop breaking.

What Comes Next

If I had to predict where this goes from here, I’d watch a few things closely.

First, watch how many Fortune 500 companies sign up for DeployCo in the next six months. The initial partners include BBVA, Goldman Sachs, and a consortium of PE firms. But real validation will come when non-partner companies start writing checks.

Second, watch Anthropic’s response. They’re moving just as aggressively, and their enterprise penetration is growing faster than any other provider. If both companies go down this path simultaneously, it could accelerate the entire market’s shift toward services-led AI adoption.

Third, watch the talent war. There are only so many engineers who can do this work. OpenAI is hiring aggressively, but so is everyone else. The salaries for forward-deployed AI engineers are already astronomical, and they’re only going up.

And finally, watch the Indian IT sector. If DeployCo proves that AI-native services can deliver better outcomes faster than traditional offshore models, the next five years could see a massive restructuring of the global IT services industry.

The Bottom Line

Here’s what you need to understand. OpenAI isn’t just launching a new business unit. They’re fundamentally reimagining what kind of company they want to be.

For the past three years, they’ve been a model company that also sells software. Going forward, they want to be a service company that also builds the best models. That’s a different identity entirely. It changes who they compete with, how they make money, and how they measure success.

DeployCo is the most honest acknowledgment yet that model excellence alone doesn’t win enterprise customers. What wins is the ability to walk into a messy, complicated, human organization and make AI work anyway.

The era of selling shovels is ending. The era of getting paid for striking gold is just beginning. And OpenAI just placed the biggest bet yet on which side of that transition will matter more.

---

The views expressed in this article are solely those of the author and do not constitute investment advice. Market conditions and company strategies remain subject to change.

April 19, 2026. Beijing Yizhuang. The morning sun was just cutting through the mist.

The starting gun fired, and 12,000 human runners and over 100 humanoid robots burst off the line together. A 21.0975-kilometer half marathon, humans and machines side by side — a scene that would’ve been pure sci-fi two years ago, now very much real.

But the result? That’s what made everyone’s jaw drop.

A robot named “Lightning” crossed the finish line with a net time of 50 minutes and 26 seconds. Let that sink in. Just over a month earlier, Uganda’s Jacob Kiplimo set the men’s half marathon world record in Lisbon at 57:20. Lightning beat it by nearly seven minutes.

One widely shared quip from a reporter on the scene: “I barely got half a sentence out, and it was already gone.”

Even more brutal? The human race winner that day, Zhao Haijie, was so stunned he said afterwards he wanted to “learn from the robots.”

It was probably the first time human runners got out-hustled this badly on a race track.

A year ago, they were basically baby walkers

If you’d watched the very first Yizhuang robot half marathon back in 2025, you probably couldn’t stop laughing.

That year, 20 teams entered. Only six robots actually finished. The champion, TianGong Ultra, clocked in at 2 hours, 40 minutes and 42 seconds. Most of the robots wobbled and stumbled all over the course — one would be running, then suddenly trip and faceplant. Another swayed side to side like it’d had one too many. Internet commenters didn’t hold back: “This isn’t a marathon, it’s a baby walker convention.”

A lot of robots couldn’t even “walk” properly. The tiny robot “Xiaopai,” made by a team called Gaoqing Power, managed just 100 meters before quitting.

Fast forward one year, and those tottering metal toddlers pulled off a mind-blowing comeback.

This year, the field ballooned nearly five times — over 100 teams, more than 300 robots. The finish rate jumped from under 30% last year to over 45%. And get this: all three podium finishers smashed the human half marathon world record. Xiaopai? It went from 100 meters last year to nearly 10 kilometers this time.

The champion’s time was slashed from 160 minutes to 50 minutes — a 110-minute improvement in a single year. That kind of evolution speed is the thing that truly sends shivers down your spine.

So how did it get this fast?

Honestly, pulling off a 6-meter-per-second sprint for 21 kilometers with a humanoid robot comes down to a three-way breakthrough: training algorithms, hardware, and perception and decision-making.

The training algorithm bit works a lot like using a cheat code in a video game. Engineers build a 3D virtual course and let the robot make infinite mistakes inside a simulator. Reinforcement learning figures out the best strategies automatically, then they deploy those strategies onto the real machine for fine-tuning. It’s like doing ten thousand mental run-throughs before the real race — no wonder it ran with so much confidence.

The hardware side is where things get really tough under the hood. Running 21 clicks means motors are under crazy load the entire time, and heat buildup in the joints is enemy number one. Lightning’s biggest secret weapon? They took the liquid cooling tech from Honor smartphones and stuck it into a robot. That kept the core motors running cool even after nearly an hour of high-intensity racing.

The joint motors themselves are top-shelf too — peak torque of 400 N·m per joint, on par with a road-going SUV. Throw in those 0.95-meter bionic long legs, and you’ve got a machine that chews up ground naturally.

On perception and decision-making, the big headline this year was autonomous navigation at scale. Last year, most robots still needed a human holding a remote control, trailing behind like somebody herding a tin kid. This year, nearly 40 percent of the robots ran fully autonomously — they saw the road, made decisions, and ran all by themselves.

Lightning had two LiDAR sensors, top and bottom, plus a satellite antenna on its head, giving it centimeter-level positioning. When it hit weird road surfaces or changing weather, it figured out the best route on its own. One of the most memorable moments: a robot reached a turning point, suddenly stopped, swayed left, swayed right, like it was genuinely hesitating about which way to go. Three or four seconds later, it turned and kept running — no remote control, no engineers anywhere near it.

What happens when robots crash?

This isn’t a closed lab track, so surprises happen. Lightning got sideswiped during the race. But using its onboard dynamic balance algorithms, it pulled itself back upright, rejoined the race, and did the whole thing with zero human intervention.

Another moment that stuck with people: a remote-controlled version of Lightning fell flat on its face during the final 100-meter sprint. Usually a robot falling is a real headache. This one got itself up in 30 seconds and kept running to the finish.

Self-correction, self-recovery — stuff that sounds basic is actually the result of a year’s worth of blood, sweat, and tears from engineers working on motion control and dynamic stability.

From making phones to making robots — what’s Honor thinking?

The company behind Lightning might already be in your pocket. It’s Honor.

A phone maker, suddenly dabbling in humanoid robots. What’s the deal?

This ties back to Honor’s Alpha Strategy. As early as MWC 2025 in Barcelona, Honor made it clear they’re transforming from a smartphone manufacturer into a global AI terminal ecosystem company, committing over 10 billion US dollars over five years in three steps.

In plain English: Honor doesn’t want to just be “the phone guys” anymore. They’re playing for the bigger AI ecosystem game.

At the 2026 Beijing Auto Show, Honor brought Lightning, another little robot called “Yuanqi Zai,” and the first new species product from the Alpha Strategy — the Robot Phone — all onto their booth. What’s a Robot Phone? Think of it as a microscopic robotic gimbal system baked into a phone body, with an AI agent brain at its core, bridging the gap between phone intelligence and robot intelligence. The phone is no longer “that boring black slab.” It now has a body that moves and senses — it tracks you, catches your smile, responds to your blink.

From phones to robots, from glowing screens to the physical world, Honor is shifting from “the phone brand” to “the AI ecology brand.” Joining a half marathon wasn’t just about grabbing headlines. It was about using the race as the ultimate stress test for their tech — proving how far and how long a robot can last in messy, real-world conditions.

Faster than humans. And then what?

Some people will ask: if robots can outrun us, is the next step to “replace” human runners?

Truth is, marathons for humanoid robots have never been about sporting competition. As the race organizers keep repeating, it’s a “technology verification” platform — a way to use a real-world scenario to push technology forward.

The real endgame for humanoid robots is stepping into factories, stepping into homes, doing the dull, dangerous work that humans really don’t want to do. Today’s marathon course is tomorrow’s real world.

Fifty minutes and twenty-six seconds is a beautiful number. But that number is just the beginning. The true finish line for robots isn’t on this 21-kilometer stretch of asphalt — it’s standing at the doorstep of every household.

When a robot helps us carry heavy stuff, cares for the elderly, or steps into dangerous places on our behalf, that’s when we’ll really feel the warmth of the tech — not just the speed.

Until that day comes, let’s end this the simplest way possible: Keep running, Lightning. And humanity’s future? Keep pushing, too.