Meta's Most Efficient AI Infrastructure for Digital Intelligence in History
Mornings With Markman - February 23rd, 2026
Meta Is Spending $135 Billion on AI This Year. The New Nvidia Chips Explain Why That Might Be a Bargain.
Last Monday, Meta and Nvidia announced a multiyear partnership covering “millions” of Nvidia’s current Blackwell and upcoming Rubin GPUs, plus standalone Grace CPUs and Spectrum-X networking hardware. Analysts estimate the deal at tens of billions of dollars, making it one of the largest single infrastructure commitments in semiconductor history. Meta is the first hyperscaler to deploy Nvidia’s Grace CPUs at large scale as standalone chips, rather than paired with GPUs.
The natural reaction from investors is fear. Meta has guided 2026 capital expenditures of $115 to $135 billion, nearly double the $72.2 billion it spent in 2025. But a set of benchmark results published the day before the deal was announced changes the math on what that money actually buys.
50x More Intelligence Per Watt
On February 16, Nvidia released SemiAnalysis InferenceX data showing that its new Blackwell Ultra GB300 NVL72 system delivers up to 50x higher throughput per megawatt compared to the previous Hopper platform. That translates into 35x lower cost per token for inference workloads.
For anyone unfamiliar with the terminology: a “token” is the basic unit of text that an AI model processes. Every time you ask an AI assistant a question, or an AI agent analyzes a document, it consumes tokens. The cost per token determines whether deploying AI at massive scale is profitable or ruinous. A 35x reduction in that cost is the difference between AI being an expensive experiment and a core business function.
The Blackwell Ultra chip also provides 1.5x higher compute performance and 2x faster attention processing compared to the standard Blackwell GPU. This matters specifically for AI agents, the software that performs multi-step tasks like writing code, managing customer service interactions, or optimizing ad targeting. These agents require continuous reasoning over long stretches of information, which is exactly the kind of workload where the efficiency gains are largest.
Microsoft Azure, CoreWeave, and Oracle Cloud are already deploying GB300 NVL72 systems for these agentic workloads. Nvidia’s next platform, Rubin, is projected to deliver another 10x improvement in throughput per megawatt over Blackwell, and Meta’s deal specifically includes access to Rubin chips as they come to market.
Why This Matters for Meta Specifically
Meta isn’t a cloud company selling computing power to others. It’s an advertising business that generated $201 billion in revenue in 2025, up 22% year over year, with 3.58 billion daily active people across Facebook, Instagram, WhatsApp, and Messenger. Every dollar Meta spends on AI infrastructure feeds directly back into making its ad targeting more precise and its user experience more engaging.
The Q4 2025 results show this flywheel working. Ad impressions grew 18% while average price per ad rose 6% in the quarter. Meta doubled the GPUs used to train its GEM ad ranking model, and the result was a 3.5% lift in ad clicks on Facebook and over 1% gain in conversions on Instagram. A unified model across Instagram Feed, Stories, and Reels produced a 3% increase in conversion rates. Redistributing ads across users and sessions delivered nearly 4x the revenue impact of simply increasing ad volume in the second half of 2025.
In plain terms: Meta is using AI to show fewer but more relevant ads, and making significantly more money doing it.
The Capex Debate Misses the Denominator
The investor concern about $135 billion in annual spending is understandable. That’s more than the entire GDP of about 140 countries. But the critique treats AI infrastructure as a static cost, like building a factory that produces the same output year after year.
The reality is closer to a factory that gets 50x more productive every two years. If each new generation of chips produces dramatically more intelligence per watt, then the return on each dollar of capex is compounding, not flat. Meta’s Q4 operating margin was 41%, down from 48% a year earlier as infrastructure spending ramps. But the company has guided that 2026 operating income will exceed 2025’s level, meaning it expects revenue growth to outpace the spending increase.
The company enters 2026 with $81.6 billion in cash and marketable securities, $115.8 billion in operating cash flow generated in 2025, and Q1 2026 revenue guidance of $53.5 to $56.5 billion, which beat analyst expectations of $51.4 billion. This isn’t a company stretching to fund a speculative bet. It’s a company with a proven monetization engine, advertising to 3.58 billion people, investing in hardware that makes that engine dramatically cheaper to operate.
What to Watch
There are real risks. Meta’s Reality Labs division lost $6 billion in Q4 alone and has accumulated nearly $80 billion in total losses since late 2020. Mark Zuckerberg said he expects 2026 to be the peak year for those losses. If the AI spending doesn’t translate into visible product improvements, like a frontier model that developers actually want to use (Meta’s Llama 4 launched to a tepid response last spring), the patience of even the most bullish shareholders has limits.
The Meta-Nvidia deal also locks Meta deep into Nvidia’s ecosystem. While Meta maintains flexibility across Nvidia, AMD, and its in-house MTIA chips, the co-design work with Nvidia on Rubin systems raises switching costs significantly. That’s a risk if Nvidia stumbles on production, and a moat if the hardware performs as benchmarked.
For anyone trying to evaluate whether big tech is overspending on AI, the efficiency data matters more than the headline dollar amount. If Meta can convert each megawatt of power into 50x more ad-targeting intelligence than it could two years ago, the $135 billion price tag starts to look less like a gamble and more like a factory upgrade with a clear payback period. The question isn’t whether the spending is large. It’s whether the output is growing faster than the cost.