Meta: Able to buy hundreds of billions of computing power, but unable to retain key personnel.

Written by: Ada , Deep Tide TechFlow

Pang Ruoming left before he even had a chance to settle into his workstation at Meta.

In July 2025, Zuckerberg poached this most sought-after Chinese engineer in the field of AI infrastructure from Apple with a multi-year compensation package totaling over $200 million. Pang Ruoming was assigned to the Meta Superintelligence Lab to be responsible for building the infrastructure for next-generation AI models.

Seven months later, OpenAI poached him.

According to The Information, OpenAI launched a months-long recruitment campaign for Pang Ruoming. Although Pang told colleagues he was "very happy working at Meta," he ultimately chose to leave. Bloomberg reported that his compensation package at Meta was tied to milestones, and leaving early meant giving up most of his unvested stock options.

$200 million can't buy seven months of loyalty.

This is not a simple story of job hopping.

One person's departure signals a whole group.

Pang Ruoming was not the first to leave.

Last week, Mat Velloso, product lead for the developer platform at Meta's Superintelligence Labs, also announced his departure. He joined Meta last July from Google DeepMind and stayed for less than eight months. Going back further, in November 2025, Turing Award winner and chief AI scientist Yann LeCun, who had been with Meta for 12 years, announced his departure to start his own business, working on the "world model" he had long championed. Russ Salakhutdinov, a key protégé of Geoffrey Hinton and vice president of generative AI research at Meta, also recently announced his departure.

To understand the talent drain at Meta AI, we must first understand just how damaging Llama 4 was.

In April 2025, Meta made a high-profile release of the Scout and Maverick models in the Llama 4 series. The official specifications were impressive, claiming that they completely outperformed the GPT-4.5 and Claude Sonnet 3.7 in core benchmark tests such as MATH-500 and GPQA Diamond.

However, this flagship model, which embodies Meta's ambitions, quickly revealed its true colors in independent blind tests conducted by third parties in the open-source community, with its actual generalization and inference capabilities falling far short of the advertised performance. Faced with strong criticism from the community, Chief AI Scientist Yann LeCun eventually admitted that the team "used different model versions to run different test sets during the testing phase in order to optimize the final score."

In the rigorous AI academic and engineering communities, this crossed an unforgivable red line. In other words, the team trained Llama 4 into a "small-town test-taker" who could only solve past exam questions, rather than a truly advanced "top student" with cutting-edge intelligence. It's like showing you a math exam paper and a programming exam paper—each individual test seems strong, but they are not actually the same model.

In AI academia, this is called "cherry picking," while in exam-oriented education, it's called "test-taking on behalf of others."

For Meta, which has always touted itself as a "beacon of open source," this turmoil directly destroyed its most valuable asset of trust within the developer ecosystem. Its immediate cost was that Zuckerberg "completely lost faith" in the engineering fundamentals of the original GenAI team, thus setting the stage for the subsequent appointment of high-ranking executives and the sidelining of core infrastructure departments.

He spent $14.3 billion to $15 billion to acquire a 49% stake in data labeling company Scale AI, bringing in 28-year-old Scale AI CEO Alexandr Wang as Meta's Chief AI Officer and establishing the Meta Superintelligence Lab (MSL). Turing Award winner LeCun was required to report to this 28-year-old in the new structure. In October, Meta laid off approximately 600 jobs at MSL, including members of the FAIR research division that LeCun had founded.

The flagship model Llama 4 Behemoth, originally planned for release in the summer of 2025, has been repeatedly delayed, from summer to fall, and finally put on hold indefinitely.

Meta has shifted its focus to developing a next-generation text model codenamed "Avocado" and an image/video model codenamed "Mango." Avocado is reportedly designed to compete with GPT-5 and Gemini 3 Ultra. Originally scheduled for release at the end of 2025, it has been delayed to the first quarter of 2026 due to unsatisfactory performance testing and training optimization. Meta is considering releasing it as closed source, abandoning the open-source tradition of the Llama series.

Meta made two fatal mistakes with its AI models. First, it fabricated benchmark data, which directly destroyed the trust of the developer community. Second, it crammed a fundamental research department like Fair, which requires a decade of dedicated work, into a product organization focused on quarterly KPIs. These two actions combined are the root cause of its current talent drain.

Self-developed chips: The other broken leg

Talent is leaving, and there are problems with the chips.

According to The Information, Meta canceled its most advanced AI training chip project, which was under development internally, last week.

Meta's self-developed chip project is called MTIA (Meta Training and Inference Accelerator). The company's initial roadmap is ambitious: MTIA v4, codenamed "Santa Barbara," v5, codenamed "Olympus," and v6, codenamed "Universal Core," are planned for delivery between 2026 and 2028. Among them, Olympus is designed to be Meta's first chip based on a 2nm chiplet architecture, aiming to simultaneously cover high-end model training and real-time inference, and ultimately replace NVIDIA's role in Meta's training cluster.

Now, this state-of-the-art training chip has been scrapped.

Meta has not been without progress; MTIA has achieved some success in inference. The MTIA v3 inference chip, codenamed "Iris," has been deployed on a large scale in Meta's data centers, primarily for Facebook Reels and Instagram's recommendation systems, reportedly reducing total cost of ownership by 40% to 44%. However, inference and training are two different things. Inference runs the model, while training practices it. Meta can make its own inference chips, but it cannot create a training chip that can directly compete with Nvidia.

This is not the first time in history. In 2022, Meta attempted to develop its own inference chip, but abandoned the project after failing in a small-scale deployment and instead placed a large order with Nvidia.

The setback in developing its own chips directly accelerated Meta's spree of outsourcing.

$135 billion in panic buying

In January 2026, Meta announced its capital expenditure budget for the year was between $115 billion and $135 billion, almost double last year's $72.2 billion. The bulk of this money will be spent on chips.

Within 10 days, three major orders were successfully placed:

On February 17, Meta signed a multi-year, cross-generational strategic cooperation agreement with NVIDIA. Meta will deploy "millions" of NVIDIA Blackwell and next-generation Vera Rubin GPUs, plus Grace discrete CPUs. Analysts estimate the deal to be worth tens of billions of dollars, making Meta the world's first supercomputing customer to deploy NVIDIA Grace discrete CPUs on a large scale.

On February 24, Meta and AMD signed a multi-year chip agreement worth $60 billion to $100 billion. Meta will purchase AMD's latest MI450 series GPUs and sixth-generation EPYC CPUs. As part of the deal, AMD issued warrants to Meta for up to 160 million ordinary shares, representing approximately 10% of AMD's shares, vesting in tranches at $0.01 per share based on delivery milestones.

On February 26, The Information reported that Meta signed a multi-billion dollar multi-year agreement with Google to lease TPU chips from Google Cloud to train and run its next-generation large language models. The two companies are also discussing Meta purchasing TPUs directly for deployment in its own data centers starting in 2027.

A social media company placed orders with three chip suppliers within 10 days, potentially totaling more than $100 billion.

This is not diversification. This is panic buying.

The Three-Layer Logic of Computing Power Anxiety

Why is Meta in such a hurry?

First, self-developed chips are no longer a viable option. The cancellation of the most advanced training chip project means that Meta will have to rely on external purchases to meet its AI training needs for the foreseeable future. While the MTIA chip for inference can handle mature applications like recommendation systems, training cutting-edge models like Avocado, which rivals GPT-5, requires NVIDIA or equivalent hardware.

Second, competitors won't wait. OpenAI has already secured massive resources from Microsoft, SoftBank, and the UAE sovereign wealth fund. Anthropic has secured supplies of 1 million TPUs and Trainium chips each from Google and Amazon. Google's Gemini 3 was trained entirely on TPUs. If Meta can't obtain sufficient computing power, it won't even be able to secure its entry into the race.

Third, and perhaps most fundamentally, Zuckerberg needs to use "purchasing power" to compensate for the lack of "R&D capabilities." The Llama 4 debacle, the loss of key talent, and setbacks in self-developed chips—these three events combined have made Meta's AI narrative fragile in the eyes of Wall Street. Signing major deals with Nvidia, AMD, and Google at this moment sends at least one signal: We have the money, we are buying, and we haven't given up.

Meta's current strategy is to invest in hardware if they can't solve the software problems, and buy chips if they can't retain talent. But the AI ​​race isn't a game you can win by simply writing checks. Computing power is a necessary condition, but not a sufficient one. Without a top-tier model team and a clear technical roadmap, no amount of chips will change anything but expensive inventory in a warehouse.

Buyer's Dilemma

Looking back at Meta's three transactions in February, an interesting detail has been overlooked by most people.

Meta bought the current Blackwell and the future Vera Rubin from Nvidia; in its deal with AMD, it bought the MI450 and the future MI455X; and it leased the current Ironwood TPU from Google, with plans to buy it directly next year.

Three suppliers, three completely different hardware architectures and software ecosystems.

This means that Meta will have to navigate between three completely different underlying ecosystems: NVIDIA's CUDA, AMD's ROCm, and Google's XLA/JAX. While a multi-vendor strategy can mitigate supply chain risks and reduce hardware procurement premiums, it will also lead to an exponential increase in engineering complexity.

This is precisely Meta's most fatal weakness. To enable a model with trillions of parameters to be trained efficiently on these three completely different underlying programming models on different hardware, it requires not only engineers who understand CUDA, but also architects who can build a cross-platform training framework from scratch.

There are probably no more than 100 people like this in the world. Pang Ruoming is one of them.

Spending $100 billion to acquire the world's most complex hardware portfolio while simultaneously losing the brains that can control it—this is the most surreal aspect of Zuckerberg's gamble.

Zuckerberg's bet

Zooming out, Zuckerberg's approach to AI over the past 18 months bears a striking resemblance to his all-in strategy of exploring the metaverse years ago:

Seeing a trend, they invest heavily and recruit a large number of people; when they encounter setbacks, they make a sudden strategic shift and invest heavily again.

The period from 2021 to 2023 was the metaverse, which resulted in losses of tens of billions of dollars each year, and the stock price eventually fell from $380 to $88. The period from 2024 to 2026 was AI, which also involved spending money without regard to cost, frequent organizational restructuring, and the same narrative of "Trust me, I have vision."

The difference is that this AI trend is indeed much more tangible than the metaverse. Meta, on the other hand, has plenty of cash to burn; its advertising business generates substantial cash flow. In the fourth quarter of 2025, Meta's revenue reached $59.9 billion, a year-on-year increase of 24%.

The problem is: money can buy chips, computing power, and even people sitting at workstations, but it can't buy people who stay.

Pang Ruoming chose OpenAI, Russ Salakhutdinov chose to leave, and LeCun chose to start his own business.

Zuckerberg's current bet is that as long as he buys enough chips, builds enough data centers, and spends enough money, he can eventually find or train people who can use these resources.

This bet might hold true. After all, Meta is one of the world's wealthiest tech companies, with over $100 billion in operating cash flow being its strongest competitive advantage. From OpenAI to Anthropic, from Google to other competitors, Meta has been continuously poaching talent. According to Qubit, nearly 40% of the 44 members on Meta's Superintelligence team came from OpenAI.

However, the cruel reality of the AI ​​race is that computing power reserves, talent lists, and model performance are all public information. The Llama 4 benchmark fraud incident proves that in this industry, you cannot maintain your lead by relying on PPT presentations and public relations.

Ultimately, the market only recognizes one thing: how good your model is.

Position in the food chain

As the AI ​​arms race enters 2026, the order of the food chain has begun to become clear:

At the top are OpenAI and Google. OpenAI boasts the strongest models, the largest user base, and the most aggressive funding. Google has a complete vertical integration of its own chips, models, and cloud infrastructure. Anthropic follows closely behind, firmly holding its position in the first tier thanks to the product strength of its Claude model and the dual computing power supply from Google and Amazon.

Meta has spent the most money, signed the most chip contracts, and reorganized the most frequently, but so far, it has not come up with a cutting-edge model that can convince the market.

Meta's AI story is somewhat similar to Yahoo's in 2005. Back then, Yahoo was one of the wealthiest companies on the internet, aggressively acquiring and spending money, but it just couldn't create a search engine like Google. Money isn't everything. Zuckerberg needs to figure out exactly what Meta wants to do with AI, instead of just buying whatever's trending.

Of course, it's too early to write Meta's obituary. 3.58 billion monthly active users, $59.9 billion in quarterly revenue, and the world's largest social dataset are assets that no competitor can easily replicate.

If the next-generation model, codenamed Avocado, can be delivered as scheduled in 2026 and return to the top tier, all of Zuckerberg's spending and restructuring will be packaged as a "strategic boldness to turn the tide." But if it falls short of expectations again, then the $135 billion will only result in rows of heated silicon wafer warehouses.

After all, Silicon Valley's AI arms race has never lacked super buyers waving their checks. What it lacks are people who know how to use that computing power to forge the future.