6,000 people, $2.5 billion.

On July 2, Microsoft announced Microsoft Frontier Company — a brand-new operating entity dedicated exclusively to enterprise AI deployment. 6,000 industry and engineering experts, $2.5 billion in initial investment, with launch partners including London Stock Exchange Group, Unilever, Land O'Lakes, and Accenture. Microsoft's commercial business CEO Judson Althoff was careful to emphasize: this “goes beyond so-called Forward-Deployed Engineering (FDE)” and would be the largest, most outcome-driven engineering organization in the industry.

But everyone could see it for what it was: the FDE model. Two days earlier on June 30, AWS had just announced a $1 billion investment to launch its own AI deployment organization. In May, OpenAI and Anthropic each partnered with private equity firms to launch similar AI services joint ventures. Four tech giants collectively betting on the same model within 60 days — that's no coincidence.

Zuckerberg's Internal Admission: Agents Aren't Progressing as Fast as Expected

Also on July 2, another story provided a critical footnote for understanding this FDE wave. According to Reuters, Zuckerberg admitted at an internal all-hands meeting on Thursday that AI Agents haven't been “accelerating in the way executives had previously expected.”

Earlier this year, Meta laid off approximately 8,000 people (~10% of its corporate workforce) and reallocated another 7,000 to AI-related teams, including a division called “Agent Transformation.” At the meeting, Zuckerberg acknowledged that these layoffs hadn't been as “clean” as they should have been, and the expected benefits from the new AI-focused company structure “haven't materialized yet.” He anticipated the company would start seeing improvements from AI investments within the next 3 to 6 months.

This statement carried the weight of a bomb. Meta's AI infrastructure spending this year is projected to reach $115-135 billion, making it one of the most aggressive players in the global AI arms race. Engineers poached to Meta Superintelligence Labs had previously complained to the press that it was a “soul-crushing gulag,” and now the CEO himself admitted progress wasn't going smoothly.

More ironically, just one day earlier on July 1, Bloomberg reported that Meta was building out a cloud business to sell excess AI compute and model access. The news sent Meta stock surging over 10%, because the market finally saw a clear path to monetizing massive investments — even if that path wasn't building the best AI models, but selling picks and shovels.

The Rise of FDE: AI Deployment Can't Rely on APIs Alone

What is Forward-Deployed Engineering? Simply put: instead of selling an API key and walking away, you send engineers directly to client sites to work alongside their teams to actually implement AI products and solve real problems. The model was made famous by Palantir and is now standard practice for AI giants.

Why is every Big Tech company suddenly doing FDE? Because two years of practice has proven one thing: improvements in AI model capability do not automatically translate into enterprise productivity gains. Buying a GPT or Claude API is easy; but truly integrating LLMs into enterprise workflows, data pipelines, permission systems, and business processes requires massive custom development, debugging, training, and ongoing maintenance. This grunt work isn't something model vendors can solve with a blog post demo.

GitLab survey data released this week corroborates this indirectly: AI tools have indeed accelerated coding speed, but overall software delivery efficiency hasn't improved. The bottleneck isn't code-writing speed — it's requirements understanding, system integration, testing, deployment, and operations — the “last mile.” That's precisely what FDE engineers are there to solve.

Microsoft's advantage is obvious: it already has engineers stationed at most Fortune 500 companies; the customer relationships are already there. Microsoft Frontier Company can kick off directly from existing customer AI deployment needs. But that also means Microsoft has conceded something: Copilot subscriptions and Azure OpenAI APIs alone aren't enough to close the enterprise AI deployment loop. Someone has to get their hands dirty and go deep into the client's business.

Do the Pick-and-Shovel Sellers Profit Before the Gold Miners?

Meta pivoting to selling compute is an even more telling signal.

As of Q1, Meta has committed $182.9 billion to AI infrastructure over the coming years. A 2,250-acre hyperscale campus in Louisiana, a “Manhattan-sized” gigawatt data center in Ohio — these investments were originally meant to support Meta's own AI superintelligence strategy and the Llama model ecosystem. But the reality is: neither Meta AI nor Llama has generated significant external revenue, and the Muse Spark model has been delayed repeatedly.

So Meta chose a pragmatic path: follow SpaceX/xAI's lead and sell off excess compute. In May, SpaceX signed over all compute at its Colossus 1 data center to Anthropic; Bloomberg Intelligence estimated the deal could generate over $50 billion by 2028, potentially reaching $100 billion by 2030. SpaceX subsequently signed similar lease agreements with Google and Reflection AI.

The CoreWeave model is being replicated by the giants: build compute infrastructure far beyond current demand, bet on future demand growth, then lease out excess capacity in the near term to defray costs. This isn't a new business model, but it reveals a structural shift in the AI industry: compute scarcity is shifting from ‘shortage’ to ‘structural surplus,’ and those who control compute are gaining leverage.

After the announcement, CoreWeave and Nebius shares fell 10.8% and 12.4% respectively — the market fears Meta's entry as a new player will make life harder for these “new cloud providers.” D.A. Davidson analyst Gil Luria didn't mince words: Meta's move shows it's “abandoning the frontier AI race” in favor of monetizing compute. But Jefferies analyst Brent Thill saw it differently: Meta is “not exiting the AI race, but converting early aggressive capacity investments into a strategic value creation path.”

Endgame: AI Enters the “Heavy Asset + Heavy Labor” Industrial Phase

Pull this week's stories together and a clear industry inflection point emerges.

Phase One (2023-2025) was the model race: whoever could train the biggest, most capable model was king. OpenAI led with GPT-4, Anthropic chased with Claude, Google and Meta followed close behind. The core narrative of this phase was “Scaling Laws” — more parameters, more data, more capability.

Phase Two (2025-H1 2026) was the Agent race: whoever could make models truly complete complex tasks autonomously would define the next interaction paradigm. But Zuckerberg's admission shows Agent progress is slower than everyone expected. Models can score high on benchmarks, but in real enterprise environments, an Agent that can reliably execute multi-step tasks remains out of reach.

We are now entering Phase Three: the deployment race. The winners of this phase won't necessarily have the strongest models — they'll be the ones who can truly embed AI into enterprise workflows and deliver measurable business value. The core competitiveness of this phase isn't SOTA scores; it's industry know-how, system integration capability, customer relationships, and ongoing service capability. That's why Microsoft, Amazon, OpenAI, and Anthropic are collectively investing in FDE — they've realized AI deployment is a labor-heavy, asset-heavy, long-haul business, not something an API call can solve.

Meta's choice represents an alternative path: if you can't build the most profitable models, be the most profitable infrastructure provider. Even $182.9 billion in compute investment, if you can't use it all yourself, makes for substantial revenue when leased out. This follows the same logic as Levi Strauss selling jeans during the 19th-century gold rush — the gold miners don't always get rich, but the pick-and-shovel sellers reliably profit.

For AI industry practitioners, this sends clear signals: the dividends from pure model optimization are narrowing; engineers and product managers who understand industries and can deliver real implementations will become more valuable; compute costs will continue falling but won't reach free; teams that master compute scheduling and optimization hold structural advantages; AI startup opportunities don't lie in building a ‘better GPT,’ but in deeply applying AI within a specific vertical.

Zuckerberg says AI investment improvements will show in 3-6 months. Microsoft says Frontier Company is the industry's largest deployment organization. Meta says “practically every week” companies proactively reach out to buy compute. AI's fantasy phase is ending; industrial deployment has begun. This phase won't have as many sexy demos or stunning benchmarks, but it's the battlefield that will truly determine how far AI can go.