The AI labs just went full Palantir. Let me explain.

For years, Palantir confused Silicon Valley because it violated the conventional rules of software. The company relied heavily on engineers deployed directly into customer organizations — working alongside governments, banks, manufacturers, and defense agencies to integrate systems, customize deployments, and adapt the software around operational workflows. To many venture capital investors, it looked inefficient. Too services-heavy. Too dependent on humans.

But Palantir was solving a deeper problem than software distribution. It was solving organizational adoption.

The company understood that in large institutions, the hardest part is rarely the technology itself. It is embedding the technology deeply enough that the organization begins to depend on it operationally. You cannot rip it out without rebuilding the way your company actually works.

That model became known as the forward-deployed engineer model.

In the past sixteen days, the two most important AI labs in the world adopted the entire playbook — and put it on private equity steroids.

—  —  —

For most of the AI boom, the dominant assumption was that the model labs would remain infrastructure providers.

OpenAI, Anthropic, and Google would build the foundational intelligence layer. Everyone else — startups, consultants, systems integrators, enterprise software vendors — would build the applications, workflows, and customer relationships on top.

That assumption just blew up — because the labs have decided that the most valuable position in the stack is not merely owning the model.

It is owning the implementation layer around the model. The deployment. The workflow integration. The operational dependency. And ultimately, the customer relationship itself.

That is what made the announcements of the last two weeks so important.

OpenAI launched DeployCo, a $10 billion joint venture backed by TPG, Bain Capital, Brookfield, Advent and a consortium of private equity firms. And here’s the wild part: OpenAI guaranteed those PE backers a 17.5% annual return for five years. That’s not a venture investment. That’s structured credit with optionality. To staff it, OpenAI acquired Tomoro, a London AI consulting firm whose clients already include Mattel, Red Bull, Tesco, and Virgin Atlantic.

Anthropic assembled its own $1.5B venture with Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, General Atlantic, GIC, Leonard Green and others.

And then the consulting industry picked sides. Bain & Company, McKinsey and Capgemini became official OpenAI partners. KPMG expanded its relationship with Anthropic and embedded Claude across its internal platform. PwC deepened its own alliance shortly after. Accenture committed billions toward enterprise AI implementation and workforce retraining around Anthropic’s models.

The labs are no longer acting like model providers.

They are becoming operational infrastructure providers embedded directly inside enterprises themselves.

—  —  —

The PE structure is the genius part. Private equity firms collectively own thousands of mid-market enterprises — companies with $50M to $5B in revenue, the exact ACV sweet spot for AI distribution. Blackstone alone has 250+ portfolio companies. TPG has 600+. OpenAI’s DeployCo partners cover 2,000+ businesses combined.

Those businesses don’t need to be convinced to take a sales call. Their PE shareholders pre-committed at the fund level. That compresses the enterprise sales cycle from twelve months to maybe twelve weeks.

And because the labs are sending forward-deployed engineers — their own engineers, with privileged access to the model roadmap — they’re not just selling a product. They’re installing the operating system of the company.

Once that installation happens, switching costs become Palantir-level. You can’t swap Claude for GPT-next for Gemini if Claude has been embedded into your tax workflow, your AML system, your client delivery platform, and your cybersecurity stack by engineers who designed it around your specific operations.

Someone on Twitter called this “Palantir-coded.” But it’s actually more aggressive than Palantir, because the labs have something Palantir wasn’t designed around: a frontier model layer that improves rapidly across every deployment. Each engagement teaches the labs which workflows matter most, where models fall short, and what to prioritize on the roadmap — feedback that flows back into the next model release.

—  —  —

The obvious casualty list is global IT services. Accenture, Deloitte, IBM Consulting, Cognizant, Infosys, Tata Consultancy Services, Wipro, HCLTech. Companies that collectively employ over three million people and generate roughly $250B in annual revenue. Their entire business model was monetizing the gap between “the software exists” and “the software is working inside your operations.” The nightmare scenario for the industry is that OpenAI and Anthropic become the control layer for enterprise operations while services firms are reduced to commoditised implementation arms.

That’s the obvious loser. The less obvious one is roughly 4,000 vertical AI startups in the US alone right now chasing mid-market enterprise customers. They raised at frothy valuations in 2024 and 2025 on the thesis that vertical AI was the obvious post-foundation-model bet.

Most of them have not yet realized that their target customer just got pre-sold to a competitor.

In the extreme scenario, when DeployCo’s forward-deployed engineers walk into a TPG portfolio company next quarter, they will not just integrate the latest model. They will offer to handle the company’s entire AI roadmap. They will identify use cases the customer hadn’t thought about. They will recommend not buying the vertical AI tool the customer was about to procure — because OpenAI can build that capability natively, the engineer is already on-site, and PE backers are putting pressure to implement AI across operations as quickly as possible.

The 2026 vertical AI graveyard is going to be much bigger than people expect.

But that does not mean the application layer disappears.

It means the shallow layer disappears.

—  —  —

The Twitter debate about whether horizontal models eat the application layer has a real answer, and it’s not the binary one most takes assume. The horizontal labs eat the shallow layer. The deep layer survives — and may actually compound faster, because the shallow competitors are getting cleared out at the same time.

The application-layer companies that lived on prompts will get eaten. GPT-for-lawyers. GPT-for-doctors. GPT-for-recruiters. Companies whose entire value proposition was a clever UI on top of someone else’s model. Sam Altman warned about this two years ago when he said OpenAI would “steamroll” pure wrappers. He was right.

The application-layer companies that own the customer relationship and the workflow data will survive. Cursor survives — its IDE integration is years of developer muscle memory the labs can’t ship in a release note. Harvey survives — its law firm deployments and document corpora compound only inside its product. Abridge survives — its Epic integration and 250+ health system rollouts are the moat, not its model. Enter survives — 300,000 lawsuits a year flowing through EnterOS is a data set the labs will never sit close enough to Brazil to build.

These companies aren’t wrappers — they’re vertical operating systems with proprietary data loops that compound only inside their product. The labs cannot replicate that without spending years inside each vertical, and there are too many verticals to cover.

—  —  —

The pipe still doesn’t go everywhere. Definitely not in large emerging markets like Brazil.

In the US, PE-owned companies employ 13 million people and contribute $2 trillion to GDP. In Brazil, the equivalent number is a rounding error.

Most of Brazil’s enterprise economy sits outside PE ownership — inside family conglomerates (Itaúsa, Votorantim, Klabin, Globo, Suzano), state-linked entities (Petrobras, Banco do Brasil), and a long tail of mid-market companies that no global PE firm has ever bought.

—  —  —

If you are building vertical AI in 2026, you are no longer competing on product. You are competing on what you own before the labs can install themselves.

Three thoughts on this.

One. Build inside a vertical so deep, so regulated, or so unloved that the horizontal labs will not follow you. Legal litigation defense. Medical billing appeals. Tax dispute resolution. Insurance claims processing. The dirtier the vertical, the sexier.

Two. Own a data loop that only compounds inside your product. Your model isn’t your moat. Your training data isn’t your moat. Your customer’s real-time workflow data, generated only when they use you, is your moat. Build that loop on day one or build something else.

Three. Look outside the US. The PE-distribution flywheel is structurally American. DeployCo’s pipe runs through TPG, Bain, and Brookfield portfolios concentrated in North America and Europe. It doesn’t run through São Paulo, Mexico City, Bangalore, Jakarta, or Lagos. The US mid-market just got more crowded. The rest of the world just got more valuable.

Which is exactly why Sequoia, Founders Fund, Ribbit, and Khosla are writing large vertical AI checks in São Paulo, Brazil.

1. KPMG just gave 276,000 employees access to Claude in a single move.

KPMG and Anthropic announced a global alliance yesterday — embedding Claude across KPMG's Digital Gateway platform and rolling it out firmwide across all 276,000 employees.

The deal also names KPMG as Anthropic's preferred consulting partner for private equity. That part matters.

The broader pattern is becoming hard to ignore.

In the past 30 days, nearly every major consulting firm has effectively aligned itself with one of the top frontier AI labs.

PwC expanded its Anthropic partnership. Deloitte effectively standardized on Anthropic at global scale — deploying Claude to 470,000 employees, creating a dedicated Claude Center of Excellence, and building industry-specific AI solutions around regulated sectors like healthcare, finance, and government. Accenture committed to a $1B+ Anthropic partnership and a 30,000-engineer training initiative. Bain, McKinsey, and Capgemini are deeply integrated into OpenAI's enterprise ecosystem.

This is not just "consulting firms adopting AI."

Consulting firms are the distribution layer for enterprise technology decisions.

When a Fortune 500 board asks: "What AI stack should we standardize around?" the answer is increasingly being shaped upstream by consultants before procurement teams even enter the conversation.

The labs are not winning enterprise customers one by one.

They are winning them in clusters — through the trusted intermediaries already embedded inside the world's largest organizations.

The underrated implication: enterprise AI adoption may consolidate far faster than most people expect — because the decision-making layer itself is consolidating.

2. Anthropic is renting $1.25B/month of compute from SpaceX.

Anthropic agreed to pay SpaceX roughly $1.25 billion every month through May 2029 for inference capacity across the Colossus 1 and Colossus 2 data centers in Memphis.

Over the life of the contract, that represents more than $40 billion in committed revenue to SpaceX.

Elon Musk called it "AI compute as a service."

This is one of the clearest signals yet that AI infrastructure is becoming a balance-sheet war.

Anthropic is effectively locking in enormous inference capacity years in advance. SpaceX is turning AI infrastructure into a new recurring revenue engine ahead of IPO — while signaling in its S-1 that it expects to sign additional large-scale compute agreements.

The industry now has a name for this model: the neocloud.

AI companies and infrastructure operators massively overbuild compute, then monetize excess capacity by acting like hyperscalers when utilization falls below internal demand.

The broader pattern is becoming clear: The AI race is no longer fought only at the model layer.

It is increasingly fought at the financing layer, the infrastructure layer, and the balance-sheet layer.

The next dominant AI companies may not simply be the ones with the smartest models.

They may be the ones that secured compute earliest, financed infrastructure most aggressively, and locked in long-term access to GPUs before the rest of the market realized scarcity was becoming structural.

$15 billion a year. From one customer. To rent compute from a rocket company.

That is what AI infrastructure looks like in 2026.

3. Nvidia just authorized another $80B in buybacks while growing Data Center revenue 92% YoY.

On the surface, this looked like another strong Nvidia quarter.

The deeper story is much more important.

Q1 FY27 revenue came in at $81.6B, up 85% year-over-year. Data Center revenue — effectively the AI number — hit $75.2B, up 92%. The company guided next quarter to roughly $91B in revenue.

At the same time, Nvidia authorized an additional $80B share repurchase program, on top of the $38.5B already remaining under prior authorization.

That gives the company nearly $120B of total buyback firepower.

It also increased its dividend 25x.

This is not normal behavior for a hypergrowth infrastructure company.

This is the kind of capital return profile you expect from Apple, Microsoft, or Saudi Aramco.

Nvidia is signaling several things simultaneously:

First: they are generating more cash than they can productively reinvest — despite the largest AI infrastructure buildout in modern technology history.

Second: management believes the AI capex cycle is durable enough to aggressively return capital while still scaling manufacturing, networking, software, and global supply chains.

Third: the entire AI stack is increasingly becoming a balance-sheet competition.

Anthropic is committing tens of billions of dollars to long-term compute agreements. SpaceX is monetizing excess GPU capacity through the "neocloud" model. Infrastructure funds are buying land and power assets. And Nvidia — the core supplier underneath the entire ecosystem — is returning nearly $120B to shareholders while still growing Data Center revenue 92%.

The picks-and-shovels analogy is becoming literal.

Nvidia is no longer behaving like a chip startup. It is behaving like the financial core of a new industrial system.

The most interesting part is that Wall Street's reaction was still mixed because some analysts expected even more.

We are now in a phase where 85% year-over-year growth — for one of the largest companies in history — is treated as slightly disappointing. That is what AI expectations have become.

~$5.4 trillion

Nvidia's market cap as of this week, after authorizing another $80B in buybacks and reporting 92% YoY Data Center growth. That's 2.5x the entire German stock market. Bigger than every publicly traded oil major on earth combined. Bigger than every silver mined in human history.

One company. Selling chips that didn't have a commercial use case at this scale four years ago. And the wildest thing is Data Center revenue went from ~$10B in FY2022 to a ~$300B ARR today. A 30x in five years.

Nuts.

Thanks for reading,

Olga 

P.S. If this issue was valuable to you please share it with a founder who needs to hear it. Let’s build LATAM’s next tech leaders—together

🎙 The J Curve  is where LATAM's boldest founders & investors come to talk real strategy, opportunity and leadership.