Smart Money Rotates from AI Models to AI Infrastructure

TL;DR

Tiger Global sold its OpenAI stake at an estimated 30-50x return, then led a $1B round in Cerebras (AI chips) the same month
Six Q1 2026 deals map to a deliberate “Autonomy Stack”: compute, orchestration, roads, airspace, homes, hotels
The thesis: AI value is moving from foundation models to the infrastructure that runs them and the applications that deploy them in the physical world
Every Q1 2026 deal puts AI into a physical domain with measurable operational metrics, not benchmark scores
Tiger returned approximately 7.9% in 2025 while competitors posted 23-44%. The thesis is high-conviction, not yet proven.

Selling OpenAI, Buying Cerebras

One of the largest AI investors in the world sold its OpenAI stake and led a $1B round in a chip company the same month. That is not a contradiction. It is a thesis.

In January 2026, Tiger Global Management sold its position in OpenAI through a secondary market transaction. The firm had invested in OpenAI in 2021 at a valuation below $16 billion. By early 2026, OpenAI’s secondary market valuation exceeded $500 billion. Tiger likely exited at 30 to 50x its entry price.

Within weeks, Tiger led a $1 billion round in Cerebras Systems at a $23 billion valuation. Cerebras builds the world’s largest AI chip: a single wafer-scale engine with 4 trillion transistors and 900,000 AI-optimized cores. Direct Nvidia competition at a fraction of the market cap.

Selling the model layer. Buying the infrastructure layer. Same thesis, different expression.

The Q1 2026 Deal Map

Tiger Global’s first quarter of 2026 was its most active investment period since the discipline reset that followed its 2022 losses. After making just nine new investments in all of 2025, the fund deployed capital into six new rounds and one extension in Q1 2026 alone. Every deal was AI-related. Every company had commercial traction, not just research promise.

The deals, ordered by conviction level:

Cerebras Systems, $1 billion at $23 billion. Tiger led this round. Their single largest check of the quarter. Cerebras had announced a $10 billion compute deal with OpenAI shortly before, covering 750 megawatts of capacity by 2028. The company’s valuation tripled in five months, from $8.1 billion in September 2025 to $23 billion in February 2026. An IPO is planned for Q2 2026 on Nasdaq with Morgan Stanley leading. Co-investors included Benchmark, Fidelity, AMD, and Coatue.

Waymo, $16 billion at $126 billion. Alphabet’s autonomous driving subsidiary was already processing over 450,000 paid rides per week. In 2025, Waymo completed 15 million rides, tripling year-over-year, with plans to expand to 20+ cities in 2026 including Tokyo and London. Tiger’s investor letter described Waymo as “the clear leader in autonomous driving” with “a product that is 10x safer than human drivers.”

Zipline, $800 million at $7.6 billion (across two tranches in January and March). The world’s largest autonomous drone delivery service: 2 million commercial deliveries, 120 million autonomous miles, operations across the US, Rwanda, Ghana, Nigeria, Kenya, and Japan. A $150 million US State Department contract for medical deliveries across five African countries. CEO target: one million deliveries per day within 2.5 years.

Mews, $300 million at $2.5 billion. A hotel property management system with an “agentic AI” roadmap. 15,000 customers across 85 countries. This was the largest hospitality software round in history.

Temporal, $300 million at $5 billion. An open-source durable execution platform. If that sounds abstract, here is the practical version: Temporal guarantees that complex AI workflows complete even when systems fail. Hundreds of thousands of developers use it. Major AI labs are customers. Tiger had led Temporal’s Series C at $1.72 billion just one year earlier.

Sunday, $165 million at $1.15 billion. A household robotics company building “Memo,” a wheeled semi-humanoid robot trained on 10 million episodes of real household routines from 500 homes. The founders, both Stanford researchers, developed Action Chunking with Transformers and Diffusion Policy, two foundational papers in robotic manipulation.

The Autonomy Stack: Six Layers of a Deliberate Thesis

These deals are not random. Map them vertically and a deliberate investment thesis emerges:

Compute layer: Cerebras provides the hardware that makes large-scale AI workloads economically viable. Tiger led this round, the strongest signal of conviction in their toolkit.

Orchestration layer: Temporal provides the software infrastructure that makes AI agents reliable in production. Durable execution means agents can run long, stateful workflows without crashing. This is the plumbing between the hardware and the applications.

Autonomous mobility: Waymo operates on roads with 450,000+ weekly paid rides.

Autonomous logistics: Zipline operates in the air with 2 million deliveries.

Autonomous home: Sunday operates in households, solving physical chores.

Autonomous vertical: Mews operates in hotels, managing operations through AI agents.

And they exited the model layer, selling OpenAI, while building this stack. The strategic read is clear: Tiger believes the value in AI is moving from the models themselves to the infrastructure that runs them and the applications that use them in physical-world settings.

Why the Model-to-Infrastructure Rotation Is Happening

Tiger’s PIP 17 investor letter, used for their latest $2.2 billion fund raise, was explicit: AI valuations are “elevated” and “in our view, at times unsupported by company fundamentals.” They used the word “humility” to describe their approach to AI investing. This from a fund that earned more money on AI than almost any other venture investor.

The argument for the rotation comes down to economics. Foundation model companies compete on a dimension, training compute, that gets more expensive every generation. OpenAI’s $110 billion raise in February 2026 (SoftBank $30 billion, Nvidia $30 billion, Amazon $50 billion) illustrates the capital intensity. These companies need enormous ongoing investment just to maintain competitive position.

Infrastructure companies face different economics. Cerebras sells hardware. Temporal sells reliability. Their revenue scales with adoption of AI broadly, regardless of which foundation model wins. They are the picks and shovels during a gold rush, a thesis as old as Levi Strauss.

Application companies in Tiger’s portfolio face yet another dynamic: defensibility through real-world operations. Waymo’s 15 million rides in 2025, Zipline’s 120 million autonomous miles, Mews’s 15,000 hotel customers across 85 countries. These are moats built from physical-world complexity, not model architecture. No amount of training compute can replicate the regulatory approvals, operational footprint, and customer trust that Waymo has accumulated over a decade of deployment.

What This Means for AI Portfolio Construction

The model-to-infrastructure rotation has implications beyond Tiger’s specific portfolio.

First, it suggests that the most sophisticated AI investors are no longer asking “which model will win?” They are asking “what infrastructure do all the models need?” This reframes the investment question entirely. You do not need to pick the model winner if you own the compute layer, the orchestration layer, and the application layer that all models depend on.

Second, Tiger’s deal pace tells a story about timing. Nine deals in all of 2025. Six new rounds plus an extension in Q1 2026 alone. This is not a fund that abandoned caution. Their PIP 17 raised the same $2.2 billion as PIP 16, a deliberate choice after PIP 15’s $12.7 billion debacle. The acceleration in deal volume while maintaining fund size discipline suggests Tiger sees Q1 2026 as a window: infrastructure layer valuations are still rational even as model layer valuations approach unsupported levels.

Third, every new deal in Q1 2026 puts AI into the physical world: hardware, roads, airspace, homes, hotels, workflow execution. Tiger’s portfolio is a bet that AI’s next chapter is not about chatbots getting better at conversation. It is about autonomous systems doing real work in real industries, with measurable operational metrics rather than benchmark scores.

The Uncomfortable Question

Tiger Global returned approximately 7.9% in 2025. D1 Capital returned 44%. Lone Pine returned 23%. Maverick returned 40%. Tiger has not recouped its high-water mark from before the 2022 crash, meaning no performance fees for over five years.

The infrastructure rotation thesis is elegant. It may also be early. Cerebras has yet to IPO. Temporal’s revenue scale is undisclosed. Sunday’s household robot is still in beta with 50 families. These are high-conviction bets on companies that are pre-public or early-commercial.

The counter-argument: Tiger’s strongest historical returns came from exactly this kind of concentrated, high-conviction positioning during periods when others were more cautious. Their early concentrated funds (PIP 1 through 10, under $3 billion each, fewer than 50 investments) generated approximately 34% gross and 23% net IRR, as reported in their investor communications. The bloated PIP 15 with 315 investments was the anomaly. What they are doing now looks more like a return to form.

Whether the infrastructure thesis pays off in 2026 or 2027 or later is the bet. But the signal for anyone watching AI investing is unambiguous: the money that made the most on foundation models is moving downstream. Following the capital means looking past the model race and into the stack that makes AI work in the real world.

Frequently Asked Questions

Why did Tiger Global sell its OpenAI position?

Tiger invested in OpenAI in 2021 at a sub-$16 billion valuation. By early 2026, the secondary market valuation exceeded $500 billion. At an estimated 30-50x return, this was a textbook profit-taking exit. Their PIP 17 letter explicitly describes AI model layer valuations as “elevated” and “at times unsupported by company fundamentals.” The exit was not a loss of conviction in AI. It was a conviction that the value is migrating from the model layer to the infrastructure layer.

What is the Autonomy Stack framework?

The Autonomy Stack is a vertical mapping of Tiger’s Q1 2026 deals: compute (Cerebras), orchestration (Temporal), autonomous mobility (Waymo), autonomous logistics (Zipline), autonomous home (Sunday), and autonomous vertical applications (Mews). Each layer is a bet on AI operating in a specific physical domain. Together they represent a thesis that AI’s value creation is moving from chatbots and benchmarks into real-world operations with measurable outcomes.

How does Tiger’s current strategy compare to its 2021-2022 approach?

The two approaches are nearly opposite. PIP 15 ($12.7 billion) made 315 investments across every category. The current strategy (PIP 16-17, $2.2 billion each) concentrates on fewer than 50 positions. The lesson from 2022 (down 56%, $42 billion in losses) was that broad spray-and-pray in tech does not survive a downturn. Concentrated, high-conviction picks with commercial traction is the reset.

Is this rotation specific to Tiger or an industry-wide trend?

Other top-tier funds are making similar moves, though with different portfolio structures. The broader signal: secondary market demand for AI infrastructure companies is rising while model-layer premiums are compressing relative to capital requirements. The infrastructure layer offers revenue that scales with AI adoption broadly, without needing to pick the model winner.

What does this mean for funds evaluating AI investments?

The Autonomy Stack framework offers one lens for portfolio construction: instead of asking “which AI model will win?”, ask “what infrastructure do all models need?” and “which real-world applications have operational moats that model advances cannot replicate?” Tiger’s answer is that physical-world complexity (regulatory approvals, logistics networks, operational track records) creates durable defensibility in ways that software moats increasingly do not.

Sources: Tiger Global PIP 16/17 investor letters (referenced by CNBC, HedgeFundAlpha), Cerebras $1B Series H (Feb 2026, Tiger-led, $23B valuation), Waymo $16B round (Feb 2026, Alphabet subsidiary), Zipline $800M Series H (Jan/Mar 2026, $7.6B valuation), Mews $300M Series D (Jan 2026, $2.5B), Temporal $300M Series D (Feb 2026, $5B), Sunday $165M Series B (Mar 2026, $1.15B), OpenAI secondary sale (Jan 2026), Bloomberg AI secondary demand analysis (Apr 2026)

Last updated: April 14, 2026

If the model-to-infrastructure rotation is shaping how your firm thinks about AI allocation, we’d welcome that conversation.

By BetterAI | We build custom AI research infrastructure for European investment firms. See how it works