2026-05-18
On May 18, two things landed at the same time.
Israeli AI company AI21 Labs announced it was laying off 60% of its workforce, shrinking from 180 to about 70 employees, ceasing independent model sales, and redirecting all resources to its Maestro agent optimization platform. That same day, Meta’s internal memo leaked: the company had reassigned 7,000 employees to AI-related roles and would begin laying off about 8,000 people (10%) on May 20. In early April, at least 1,000 top engineers had already been forcibly transferred into a new AI division, with refusal meaning termination. Meta’s 2026 capital expenditure is projected at $125-145 billion.
These two stories won’t appear in the same news article, but they form a complete picture only when placed side by side. On one side, model companies are shrinking and pivoting. On the other, a platform company is using organizational force to redirect itself toward AI. Both point to the same conclusion: the middle ground is disappearing. AI model companies that are neither hyperscale platforms nor possess unique moats are exiting independent model sales, one after another.
AI21 is not the first, and it won’t be the last.
In the 16 months from March 2024 to July 2025, five prominent independent AI model/product companies were absorbed by big tech firms using nearly identical deal structures: Inflection AI’s Mustafa Suleyman and core team joined Microsoft (March 2024), Adept AI’s founders and roughly 80% of the team joined Amazon (June 2024), Character.AI’s Noam Shazeer returned to Google (August 2024), Covariant’s co-founders joined Amazon (August 2024), and Codeium/Windsurf’s team and IP were split between Google and Cognition (July 2025).
These deals all used the same “reverse acqui-hire” structure — big companies hire the team and license the technology without directly acquiring the company, sidestepping antitrust review. The structure itself is a signal: the buyers didn’t think these companies’ standalone businesses were worth buying. They only wanted the talent and IP.
Other cases point in the same direction. Stability AI’s CEO Emad Mostaque resigned in March 2024; the company cut staff under new leadership, pivoted from open-source models to commercial licensing, and its revenue data remains uncertain. China’s 01.AI (founded by Kai-Fu Lee) completely stopped pre-training large language models in March 2025, switching to selling customized solutions based on DeepSeek’s models. Humane’s AI Pin permanently shut down after selling its assets for $116 million, far below the roughly $240 million it had raised.
AI21’s specific situation shares the same root causes as this group. According to a series of reports from Israeli financial outlet Globes, AI21 raised roughly $336 million in actual deployed capital since its 2017 founding (a fraction of OpenAI’s $63 billion and Anthropic’s $46 billion), with a peak valuation of $1.4 billion, and had already begun seeking a sale by late 2025. The company successively released Jurassic-1, Jurassic-2, and Jamba language models. In 2021, Jurassic was still a meaningful competitor to GPT-3. But by 2024-2025, after GPT-4, Claude, and Gemini had built systematic advantages across multimodality, reasoning, and long-context capabilities, the capital cost of catching up grew exponentially.
Globes reported in its own words that AI21 “has struggled to sell its language models” and “has not garnered great commercial success.” The company never publicly disclosed revenue; estimates place it in the tens of millions. By late 2025, it began seeking a sale, negotiating first with Nvidia and then with Nebius — neither round reached a deal. After the Nvidia talks collapsed in January 2026, AI21 issued a statement denying “acquisition talks,” but the wording focused narrowly on the word “acquisition” itself, suggesting Nvidia had in fact proposed something like an acqui-hire that was rejected. Nebius negotiations began in April 2026 and collapsed by May, though the two sides did sign a Maestro implementation agreement — Nebius apparently concluded that buying the product was cheaper than buying the company. After the acquisition fell through, the layoffs and pivot were announced almost simultaneously.
One sentence in the company’s statement is its own verdict: “The company will stop selling models per se — they are an important infrastructure for its AI expertise, but not a sufficient revenue source in themselves.”
That sentence summarizes the entire logic.
Meta made the opposite choice under the same logic.
According to cross-confirmed reporting from The Information, Reuters, and the NYT, Meta forcibly transferred at least 1,000 top engineers into the newly created Applied AI Engineering division inside Reality Labs in early April 2026. Internally, the move was called “the Draft,” and engineers who refused the transfer were threatened with termination. This approach is extraordinarily rare in Silicon Valley — in an industry known for talent mobility, using the threat of firing to force engineers into specific teams signals that management has placed this priority above organizational health.
The May 18 internal memo disclosed further details: the company planned to cut about 8,000 jobs (10% of the global workforce), while reassigning roughly 7,000 employees into four new AI organizations. Another 6,000 open roles were closed. HR chief Janelle Gale wrote in the memo about “AI native design principles” and a “flatter structure with smaller teams of pods/cohorts” — meaning the company’s architecture itself is being redesigned as an AI-native organizational form, stripping out management layers and replacing traditional hierarchies with small autonomous teams.
Employee reaction can be summarized through an interview from SF Standard. A decade-tenured Meta employee gave an extensive anonymous interview. Partial quotes: “I tend to cry in the shower.” “This is as anxious and stressed as I have ever been at a job.” “Even if we haven’t lost our jobs to AI yet, we’re being commoditized in advance.” The employee described a “gallows humor” culture — layoff memes and dancing skeletons circulating in chat groups that include VPs. Many employees are on mental health leave, described as an “open secret.”
Two additional factors have deepened the internal rift. First, in April, Meta launched a program tracking employees’ keystrokes, mouse movements, clicks, and screen activity to train AI models on “how people actually complete everyday tasks using computers.” CTO Andrew Bosworth explicitly stated “there is no option to opt-out on your corporate laptop”, and over 1,000 employees signed a petition opposing this privacy violation. Second, some employees began manually disabling AI note transcription during video calls so they could speak candidly about layoff rumors — effectively treating the company’s own AI tools as surveillance infrastructure.
Zuckerberg’s actions have a clear historical reference point: Google’s “Code Red,” declared after ChatGPT launched in late 2022. Sundar Pichai reassigned large numbers of engineers from Search, Assistant, and other divisions to AI teams, launching Bard (later Gemini) within weeks. Google did produce competitive AI products, but at the cost of a significant exodus of top researchers to OpenAI and Anthropic — talent that later directly became Google’s AI competitors. Meta’s current “Draft” mechanism is more aggressive than Google’s Code Red (mandatory vs. voluntary internal transfers), making the attrition risk correspondingly higher.
Model companies dying and platform companies betting their lives share the same economic logic.
Model capabilities are converging. AI benchmarking platform Artificial Analysis tracks 516 models from 51 model creators. Its Intelligence Index shows GPT-5.5, Claude, Gemini, Grok, and DeepSeek forming a tight cluster at the top, with open-weight models approaching or matching proprietary models across multiple dimensions. Gaps persist in specific dimensions — particularly reasoning and agentic tasks — but differences in general text generation quality are no longer sufficient to support an independent commercial model.
The direction of token pricing is even clearer. When GPT-4 launched in March 2023, it priced at roughly $30 per million input tokens. By late 2025, comparable capability had dropped to about $0.30 — roughly a 100x decline in two years. For any given open-weight model, Artificial Analysis tracks 21 competing inference providers — identical product, with price and speed as the only differentiating dimensions.
a16z is the most explicit articulator of this judgment. In their late-2025 Big Ideas 2026, four a16z partners named “The Enterprise Orchestration Layer” as a core annual thesis. A May 2026 follow-up piece, From System of Record to System of Intelligence, made the argument more directly: the locus of value in enterprise software is migrating upward, from the database layer (CRM, ERP) to the orchestration layer that reads that data, reasons about it, and takes action. “A foundation model is not, by itself, a GTM application, any more than Oracle’s database engine was a CRM. Between the model and the customer sits an enormous amount of unglamorous and domain-specific work — that work is the new GTM application layer.”
SaaStr’s Jason Lemkin provided a concrete micro-case: his company reduced Salesforce seats from 10+ to 2 human seats plus 1 API seat, while total spend rose from $12,000 to $22,000 annually. The model layer got cheaper; the orchestration layer captured more value.
So if the model layer is commoditizing, what kind of independent model company can still survive?
A set of moat types emerges from the counter-examples, with each surviving company holding at least one, often several.
First, sovereign/geopolitical backing. Mistral AI reached a $14 billion valuation largely not because its models are better than anyone else’s, but because it positioned itself as the standard-bearer for European AI sovereignty. ASML (the Dutch semiconductor equipment giant) invested €1.5 billion and holds an 11% stake. Mistral secured support from the French and Swedish governments for building its own data centers. Cohere received $240 million in public funding from the Canadian government for domestic AI compute infrastructure and announced a merger with Germany’s Aleph Alpha in April 2026, building a transatlantic AI sovereignty narrative. DeepSeek’s implicit Chinese government and PLA affiliations provide similar backing and domestic market access. AI21 is an Israeli company, but lacked a corresponding sovereignty narrative.
Second, distribution ownership rather than model quality. Anthropic reached a $380 billion valuation not because Claude is uniquely superior in capability, but because it is simultaneously embedded in AWS (Amazon cumulative investment $8 billion), Google Cloud ($2.5+ billion), and Microsoft Azure ($15 billion deal), and secured a $200 million U.S. Department of Defense contract. Its Model Context Protocol is becoming an industry standard for connecting AI models to external data. AI21’s models are available on AWS Bedrock, but that is a single channel, not any form of channel ownership.
Third, patient capital with non-VC incentive structures. DeepSeek’s parent company High-Flyer is a Chinese quantitative hedge fund that does not need to generate commercial revenue from its models. DeepSeek has explicitly stated it has “no immediate plans for commercialization.” Anthropic’s funding scale ($30 billion Series G) creates years of runway. Mistral’s strategic investors (ASML, Microsoft) are aligned on long-term positioning rather than near-term exits. By contrast, AI21’s $336 million in VC funding was insufficient to create patient capital — once model sales proved non-viable, the VC return logic demanded either a pivot or an exit, and both eventually happened.
Fourth, vertical specialization rather than generalist competition. Cohere reached $240 million ARR (doubling from $100 million in under a year) by abandoning the frontier model benchmark race and focusing on enterprise deployments in regulated industries (finance, healthcare, defense). Its North platform provides a secure AI workspace for enterprises that cannot use public cloud AI services. Cohere’s positioning gives it an advantage in domains where switching costs are high and compliance requirements create natural barriers to entry. AI21 never specialized — it kept trying to compete as a general-purpose model provider.
Fifth, open-source as ecosystem strategy rather than marketing. Mistral and DeepSeek release open-weight models (under Apache-2.0 and MIT licenses, respectively), creating developer community lock-in. But open-source alone is not enough — 01.AI was also open-source and still had to abandon model building. Open-source only becomes a moat when layered with other advantages (Mistral’s sovereignty backing, DeepSeek’s cost advantage).
Even companies with these moats need to make the same choice — all of them, without exception, are actively moving up (into the application layer) or down (into the infrastructure layer). Anthropic’s Claude Code is a developer tool, not a model API. Mistral’s Le Chat is a consumer product. Cohere sells solutions, not APIs. Companies that purely sell model access, regardless of scale, are proving to have a structurally unsustainable business model.
Having gone around the full circle, the conclusion is simple.
The model layer is commoditizing. This is happening simultaneously across price, capability, and capital flow — it’s not a prediction, it’s in progress. The path of selling models independently is closed. AI21’s own layoff statement effectively summarized it for the entire industry: “Models per se are not a sufficient revenue source.”
For builders, only two things remain. First, treat model selection as a runtime optimization parameter, not an architectural decision — your orchestration layer should always be multi-vendor. Second, what’s valuable lives in your orchestration layer — domain logic, accumulated user context, multi-step workflow automation. These are things no model, no matter how strong, can replace for you.