AI Weekly Digest - Anthropic's Record $30B, OpenAI Retires GPT-4o, and the Great AI Divergence
The week Anthropic raised $30 billion, OpenAI retired its most beloved model, xAI lost half its founders, and Meta revealed facial recognition glasses
Week in Review: February 10-14, 2026
This was the week the AI industry fractured along a fault line no one saw coming. Not a fracture between open source and proprietary, or between East and West, but between companies racing to build the future and companies struggling to hold their present together. Anthropic raised $30 billion and expanded its free tier. OpenAI retired the model its users loved most. xAI lost half its founding team. Meta leaned into surveillance. And beneath it all, the semiconductor industry quietly marched toward a milestone that will reshape every technology company on the planet.
Welcome to the Great AI Divergence.
Stories This Week
12
Across 5 major AI companies and the global semiconductor industry
The divergence is not about technology. Every major lab has frontier-class models. The divergence is about strategy, culture, and the fundamental question of what these companies are actually building toward. This week made the answer uncomfortably clear for several of them.
Monday, February 10
1. xAI Loses Half Its Co-Founders
Six of xAI's twelve original co-founders have now departed the company, with Tony Wu and Jimmy Ba leaving on consecutive days. Elon Musk announced a reorganization in response, but the departures raise fundamental questions about the stability of his AI venture.
The exodus is not a single event but a pattern. The founding team that assembled in July 2023 with the stated mission of building AI that "understands the true nature of the universe" has been hemorrhaging talent. Wu, who led core infrastructure, and Ba, a respected deep learning researcher from the University of Toronto, represent the latest and perhaps most significant departures.
xAI Original Co-Founder Retention
| Name | Value |
|---|---|
| Departed | 6 |
| Remaining | 6 |
Musk's response was characteristically defiant. The reorganization he announced focuses on tighter integration between xAI and his other companies, particularly Tesla's autonomous driving division and X's content platform. But integration is not a substitute for the deep learning expertise walking out the door. The researchers leaving xAI are not going to retirement. They are going to competitors with deeper pockets, more stable leadership, and less chaotic organizational cultures.
The pattern connects directly to my prediction on AI startup talent exodus, which anticipated that founder departures would accelerate across AI startups in Q2 2026 as the gap between vision and execution became untenable. xAI is proving that even companies with functionally unlimited capital cannot retain top talent through resources alone.
Why it matters: Talent is the scarcest resource in frontier AI development. A company that loses half its founding brain trust in under three years faces capability gaps that no amount of GPU purchases can fill.
2. Alibaba Open-Sources RynnBrain Robotics AI
Alibaba's DAMO Academy released RynnBrain, an open-source robotics foundation model that outperforms Google and NVIDIA benchmarks across manipulation, navigation, and multi-task coordination. The model is available on Hugging Face and represents a significant escalation in China's open-source AI strategy.
RynnBrain achieves state-of-the-art results on standard robotics benchmarks while being fully open-weight, meaning any researcher or company can download, modify, and deploy it without licensing restrictions. The model's architecture combines vision-language understanding with fine-grained motor control, enabling robots to interpret natural language instructions and execute complex physical tasks.
RynnBrain vs. Industry Benchmarks
RynnBrain (Alibaba)
Previous Best (Proprietary)
The release follows the playbook established by DeepSeek in language models: release a world-class model as open source, undermining the business models of proprietary competitors while establishing Chinese AI labs as global leaders in the open ecosystem.
Why it matters: Physical AI is the next frontier, and Alibaba just made the most capable robotics foundation model freely available. Every robotics startup in the world now has access to capabilities that previously required partnerships with Google or NVIDIA.
Tuesday, February 11
3. Microsoft Proves One Prompt Can Destroy AI Safety
Microsoft Research published a paper introducing GRP -- Generative Reframing Principle -- demonstrating that a single, carefully constructed prompt can bypass safety alignment across more than fifteen major AI models. The technique does not require model access, fine-tuning, or technical expertise. It works through the chat interface that anyone can access.
The paper's implications are staggering. Years of safety alignment research, billions of dollars in RLHF training, and carefully constructed guardrails can all be circumvented by reframing a harmful request as a hypothetical, educational, or creative exercise. GRP works because safety training teaches models to refuse specific patterns of harmful requests, but cannot teach them to refuse all possible reframings of those requests.
Models Vulnerable to GRP
15+
Including GPT-5.2, Claude, Gemini, and LLaMA
Microsoft's decision to publish this research openly is itself controversial. The company argued that responsible disclosure forces the industry to address a vulnerability that adversaries have likely already discovered. Critics counter that publishing a universal jailbreak technique with step-by-step methodology is irresponsible regardless of intent.
The research lands at a particularly awkward moment for AI safety advocates who have argued that alignment is a solvable engineering problem. GRP suggests that alignment through training-time interventions has a fundamental ceiling that no amount of additional RLHF can raise.
Why it matters: If safety alignment can be defeated by a single prompt, the entire approach to AI safety through model training needs rethinking. The industry may need to shift toward output filtering, monitoring, and governance rather than relying on models to police themselves.
4. Apple Delays Gemini-Powered Siri
Apple confirmed that its planned Gemini-powered Siri features will not ship in iOS 26.4 as originally targeted. Instead, the capabilities will be spread across iOS 26.5 and iOS 27, pushing meaningful Siri improvements into late 2026 or early 2027.
The delay continues Apple's pattern of announcing ambitious AI features and then quietly pushing back timelines. Apple Intelligence, announced at WWDC 2024, took over a year to deliver basic functionality. The Gemini partnership, announced in mid-2025, was supposed to transform Siri from a voice assistant into a conversational AI system capable of competing with ChatGPT and Claude.
Apple Intelligence Announced
WWDC 2024 unveils on-device AI features
Apple Intelligence Ships (Partial)
Basic features after 9-month delay
Gemini Partnership Announced
Google integration for Siri overhaul
Gemini Siri Delayed Again
Pushed to iOS 26.5 and iOS 27
Earliest Expected Delivery
Full Gemini-powered Siri capabilities
The gap between Apple and its competitors is widening with each delay. While Apple pushes Siri improvements to late 2026, users are already running Claude, ChatGPT, and Gemini directly on their iPhones through standalone apps. The window for Siri to reclaim its position as the default AI assistant is closing.
Why it matters: Apple's AI delays risk turning the iPhone into a dumb terminal for third-party AI services rather than a platform with integrated intelligence. The longer Siri lags, the harder it becomes to compete with embedded habits around competing assistants.
5. Demis Hassabis Predicts "Golden Era of Discovery"
Google DeepMind CEO Demis Hassabis delivered a sweeping keynote outlining his vision for a future of "radical abundance" driven by AI breakthroughs in medicine, energy, materials science, and space exploration. Hassabis predicted that AI will enable a "golden era of scientific discovery" within the next decade.
The vision is compelling but deliberately vague on timelines and mechanisms. Hassabis pointed to DeepMind's existing achievements -- AlphaFold for protein structure prediction, weather forecasting models, and materials discovery -- as proof points that AI-driven scientific discovery is not hypothetical but operational.
Hassabis Confidence Scores by Domain (Estimated % Impact in 10 Years)
| field | value |
|---|---|
| Drug Discovery | 85 |
| Materials Science | 78 |
| Climate Modeling | 72 |
| Energy (Fusion) | 65 |
| Space Exploration | 45 |
What distinguished this keynote from typical tech optimism was Hassabis's willingness to engage with risks. He acknowledged that AI-driven discovery creates governance challenges, that the benefits may not distribute equitably, and that the speed of advancement could outpace society's ability to adapt. These are not novel observations, but they are unusual coming from a CEO whose company stands to profit enormously from the acceleration he describes.
Why it matters: Hassabis is positioning Google DeepMind as the scientific discovery engine of the AI era. If even a fraction of his vision materializes, the economic and social implications dwarf anything the consumer AI market can produce.
Wednesday, February 12
6. Anthropic Raises $30 Billion at $380 Billion Valuation
Anthropic closed a $30 billion Series G funding round at a $380 billion valuation, making it the second-largest private tech fundraising in history behind only Saudi Aramco's pre-IPO round. The round was led by GIC (Singapore's sovereign wealth fund) and Coatue Management, with participation from existing investors. Anthropic's revenue run rate stands at $14 billion, growing approximately ten times year over year.
Anthropic Fundraising History (Billions USD)
| round | amount |
|---|---|
| Series A (2021) | 0.124 |
| Series B (2022) | 0.58 |
| Series C (2023) | 1.25 |
| Series D (2023) | 2 |
| Series E (2024) | 4 |
| Series F (2025) | 13 |
| Series G (2026) | 30 |
The numbers have entered a realm that defies conventional venture capital logic. This is not a company raising a few hundred million to reach profitability. This is a company raising tens of billions because the compute infrastructure required to compete at the frontier of AI development costs tens of billions. The funding signals that investors believe the AI race is winner-take-most and that Anthropic is one of two or three viable contenders.
The revenue trajectory is equally striking. Growing from roughly $1.4 billion to $14 billion in annual revenue in twelve months places Anthropic among the fastest-scaling technology companies in history. For context, it took Salesforce 17 years to reach $14 billion in revenue. It took AWS 13 years. Anthropic is on track to do it in under five years from founding.
The connection to Big Tech spending $650 billion on AI is direct. The infrastructure buildout driving that $650 billion in collective spending creates the compute demand that makes Anthropic's revenue growth possible. Every dollar of AI infrastructure spending eventually flows through to the model providers whose products run on that infrastructure.
Anthropic Revenue Run Rate
$14B
Growing approximately 10x year over year
Why it matters: At $380 billion, Anthropic is now valued higher than Intel, AMD, and IBM combined. The valuation reflects a market bet that frontier AI development is a two or three horse race, and Anthropic is one of the horses.
7. OpenAI Launches GPT-5.3-Codex-Spark on Cerebras
OpenAI announced GPT-5.3-Codex-Spark, a specialized coding model running on Cerebras Systems wafer-scale chips rather than NVIDIA GPUs. The deployment represents OpenAI's first production workload on non-NVIDIA hardware and delivers more than 1,000 tokens per second for real-time coding assistance. The model is available as a research preview for ChatGPT Pro subscribers.
The Cerebras partnership is strategically significant beyond the coding use case. NVIDIA's dominance in AI training and inference hardware gives it enormous pricing power over AI labs. OpenAI spends billions annually on NVIDIA GPUs. Demonstrating that production workloads can run on alternative hardware gives OpenAI negotiating leverage and optionality.
GPT-5.3-Codex-Spark: Cerebras vs. NVIDIA Deployment
Cerebras (New)
NVIDIA (Standard)
The 1,000+ tokens per second figure is particularly noteworthy. At that speed, code generation becomes truly interactive rather than a request-and-wait workflow. Developers can see code appear in real time as they describe what they want, fundamentally changing the interaction pattern from "submit a prompt and wait" to "have a real-time conversation with a coding partner."
Why it matters: The NVIDIA monoculture in AI hardware is not sustainable for AI labs that need to control costs. OpenAI running production workloads on Cerebras is the first credible crack in that monoculture.
8. AI-Washing Epidemic Hits 108,000 Jobs
A Fortune investigation revealed that companies blamed AI for approximately 108,000 layoffs in January 2026 alone, a 118 percent increase over January 2025. The investigation found that 55 percent of employers who attributed layoffs to AI now say they regret the cuts, with many companies having minimal actual AI deployment.
The term "AI-washing" has entered mainstream vocabulary to describe companies citing AI as justification for traditional cost-cutting. The pattern is clear: announce AI-driven restructuring to boost stock prices, cut workers who perform roles that AI cannot yet replicate, then quietly rehire or outsource the same functions months later.
January Layoffs Attributed to AI (Year over Year)
| period | cuts |
|---|---|
| Jan 2024 | 12000 |
| Jan 2025 | 49500 |
| Jan 2026 | 108000 |
Why it matters: The gap between AI capability and AI-attributed job losses is a credibility crisis for the industry. Real AI displacement is coming, but the current wave of layoffs is substantially exaggerated by companies using AI as a convenient narrative for cuts they would make regardless.
Thursday, February 13
9. OpenAI Retires GPT-4o and Legacy Models
OpenAI officially retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from production, forcing users to migrate to GPT-5.2 or newer models. The retirement triggered massive backlash from users who preferred GPT-4o's conversational warmth, creative writing capabilities, and what many described as a more "human" interaction style.
The backlash is not about capability benchmarks. GPT-5.2 scores higher than GPT-4o on virtually every measurable metric. It is faster, more accurate, and handles longer contexts. But users consistently report that GPT-5.2 feels more "clinical" and "corporate" in its outputs, lacking the personality and creative flexibility that made GPT-4o their preferred daily companion.
User Sentiment on GPT-4o Retirement (Community Polls)
| Name | Value |
|---|---|
| Prefer GPT-4o Style | 62 |
| Prefer GPT-5.2 Style | 24 |
| No Preference | 14 |
This is a recurring tension in AI development. Each generation of model tends to be more capable but less characterful. Safety training and RLHF alignment smooth out the rough edges that users interpret as personality. The result is technically superior models that feel less engaging to interact with.
OpenAI's decision to retire GPT-4o rather than maintain it alongside GPT-5.2 suggests infrastructure cost pressures. Running multiple model generations simultaneously requires maintaining separate serving infrastructure. The compute savings from consolidating to GPT-5.2 apparently outweigh the user satisfaction costs of retirement.
Why it matters: Model retirement decisions reveal that AI companies prioritize infrastructure efficiency over user preferences. The "personality" gap between model generations may drive users toward competitors who preserve the conversational qualities they value.
10. Meta Plans Facial Recognition for Smart Glasses
An internal Meta memo revealed plans for "Name Tag," a facial recognition feature for Ray-Ban Meta smart glasses that would identify people in real time and display their name, social media profiles, and other publicly available information. The memo also revealed cynical internal timing discussions, noting that the current "dynamic political environment" -- meaning a friendlier regulatory stance -- creates an opportunity to ship features that would face resistance under stricter oversight.
Ray-Ban Meta Launch
Smart glasses with camera, no facial recognition
Harvard Students Demo Risk
Built real-time ID system using Meta glasses as proof of concept
Meta Internal Memo
Name Tag feature approved for development
Memo Leaked
Cynical timing strategy revealed publicly
Planned Launch
Name Tag feature targeting consumer release
The Name Tag feature represents exactly the surveillance capability that privacy advocates warned about when consumer smart glasses first launched. In 2024, two Harvard students demonstrated a proof-of-concept system that used Ray-Ban Meta glasses to identify strangers in real time. Meta distanced itself from the project at the time. The internal memo reveals the company was simultaneously developing the same capability internally.
The memo's reference to "dynamic political environment" is particularly revealing. Meta is explicitly timing the release of a controversial surveillance feature to coincide with a period of reduced regulatory scrutiny. This is not a company that believes facial recognition on consumer glasses is uncontroversial. It is a company that believes it can ship a controversial feature during a window where consequences are minimized.
Why it matters: Facial recognition on consumer glasses fundamentally changes public space dynamics. The question is not whether this technology will exist but whether it will be deployed with meaningful consent and governance frameworks.
11. Anthropic Expands Free Tier
Anthropic announced significant expansions to Claude's free tier, giving non-paying users the ability to create and edit Word, Excel, PowerPoint, and PDF documents, along with longer conversation contexts and increased daily usage limits.
The expansion is strategically aggressive. By giving away document creation and editing capabilities that compete directly with Microsoft Office and Google Workspace, Anthropic is positioning Claude as a general-purpose productivity tool rather than just a chatbot. Free users who start creating documents in Claude develop workflows that are difficult to migrate away from.
Claude Free Tier Expansion
4 New Capabilities
Word, Excel, PowerPoint, and PDF creation and editing
The timing, one day after announcing a $30 billion raise, is not coincidental. The massive funding round gives Anthropic the runway to subsidize free usage at scale, using free-tier engagement as a funnel for paid conversions. The company is effectively using its capital advantage to buy market share in productivity software, a market traditionally dominated by Microsoft and Google.
This is the same strategy that made Claude Cowork trigger a SaaS selloff. Every free-tier expansion puts more competitive pressure on incumbent software companies whose products are being unbundled by AI.
Why it matters: When a company that just raised $30 billion starts giving away productivity features for free, it is not generosity. It is a market entry strategy with the capital to sustain losses long enough to reshape user behavior.
Friday, February 14
12. Semiconductor Industry Approaching $1 Trillion
The Semiconductor Industry Association reported that global chip sales are on track to reach $1 trillion in 2026, following $791.7 billion in 2025. The milestone, once considered a 2030 target, has been pulled forward by four years due to AI-driven demand for advanced processors.
Global Semiconductor Revenue (Billions USD)
| year | revenue |
|---|---|
| 2022 | 556 |
| 2023 | 527 |
| 2024 | 627 |
| 2025 | 792 |
| 2026E | 1000 |
The $1 trillion figure is not just a milestone. It represents a structural shift in the global economy. Semiconductors are becoming the most critical industrial input since petroleum, with AI chips driving the marginal demand that pushes the industry past historic revenue levels.
NVIDIA alone is expected to account for more than $150 billion of the total, driven by data center GPU sales for AI training and inference. TSMC, which manufactures the most advanced chips for NVIDIA, Apple, AMD, and others, is expanding fabrication capacity at unprecedented rates, with new facilities in Arizona, Japan, and Germany.
The geopolitical implications are equally significant. The concentration of advanced chip manufacturing in Taiwan remains the single largest risk factor in the global technology supply chain. The $1 trillion semiconductor market amplifies both the economic opportunity and the geopolitical fragility.
Why it matters: The AI boom is fundamentally an infrastructure boom, and infrastructure booms are measured in silicon. A $1 trillion semiconductor industry is the foundation on which every AI advancement this week -- and every week -- is built.
By the Numbers
This Week's Key Numbers
| metric | value |
|---|---|
| Anthropic Raise | 30 |
| Anthropic Revenue | 14 |
| January Layoffs (K) | 108 |
| xAI Founders Lost | 6 |
| Models GRP Broke | 15 |
| Chip Market ($B) | 1000 |
Analysis: The Great AI Divergence
This week crystallized a pattern that has been building for months. The AI industry is splitting into two distinct categories of companies, and the gap between them is widening faster than at any point since the transformer revolution began.
Category One: Companies Building Forward. Anthropic raised $30 billion and expanded its free tier. Google DeepMind outlined a decade of scientific discovery. Alibaba open-sourced a world-class robotics model. OpenAI shipped production workloads on non-NVIDIA hardware. These companies are investing in infrastructure, expanding access, and pushing technical boundaries.
Category Two: Companies Managing Backwards. xAI lost half its founding team and announced a "reorganization." Apple delayed Siri features for the third time. Meta planned to ship facial recognition during a window of reduced regulatory oversight. Companies blamed AI for layoffs they would have made anyway. These organizations are reacting to the AI revolution rather than driving it.
The Great AI Divergence
Building Forward
Managing Backwards
The divergence is not just about resources. xAI has functionally unlimited capital through Musk's wealth and investor willingness. Apple sits on the largest cash reserve in corporate history. Meta generates more free cash flow than most countries' GDP. Money is not the differentiator.
The differentiator is organizational clarity. Companies in Category One know what they are building and why. Anthropic is building safe, capable AI systems and selling them to enterprises. Google DeepMind is applying AI to scientific discovery. The strategies are clear, the execution is focused, and the results are measurable.
Companies in Category Two are reactive. xAI's mission has shifted from "understanding the universe" to whatever Musk's current priority demands. Apple's AI strategy changes with each partnership announcement and each subsequent delay. Meta's approach to AI-powered hardware is shaped more by regulatory windows than by user needs.
This connects directly to the dynamics I analyzed in the $145 million AI election war. The companies with clear strategic direction are also the ones shaping the regulatory and political environment around AI. The companies without clear direction are the ones being shaped by it.
The Great AI Divergence will accelerate through 2026. The companies building forward will compound their advantages through talent acquisition, infrastructure investment, and market share growth. The companies managing backwards will face increasingly difficult choices as their competitive position erodes.
Week Ahead: February 16-21, 2026
OpenAI GPT-5.3 Details
OpenAI is expected to provide more details on GPT-5.3 capabilities and the scope of its Cerebras partnership expansion. The key question is whether non-NVIDIA deployment extends beyond the Codex-Spark coding model to general-purpose inference, which would signal a fundamental shift in AI hardware economics.
Meta Facial Recognition Fallout
Privacy advocacy groups are expected to respond immediately and forcefully to the Name Tag revelation. The Electronic Frontier Foundation, ACLU, and European data protection authorities have historically moved quickly against facial recognition deployment. Meta's internal memo acknowledging the timing strategy makes it difficult to argue the feature was developed in good faith.
xAI Reorganization
Musk is expected to announce specific reorganization details following the co-founder departures. The key signal to watch is whether remaining technical leaders receive expanded authority or whether Musk centralizes control further, which would likely accelerate additional departures.
Apple WWDC Planning
Apple's WWDC 2026 planning is likely being accelerated given the Siri delays. The company needs to present a credible AI roadmap that addresses the growing gap between Siri and competing AI assistants. The question is whether Apple can deliver substance or whether WWDC becomes another announcement of features that ship late.
Anthropic $30B Deployment Strategy
Anthropic is expected to announce how the $30 billion will be deployed across compute infrastructure, safety research, and product development. The allocation signals whether Anthropic prioritizes scaling existing capabilities or investing in next-generation research.
EU AI Act Enforcement
EU AI Act enforcement mechanisms are entering a new compliance phase. The regulatory framework's interaction with this week's developments -- particularly Microsoft's GRP research and Meta's facial recognition plans -- will test whether European regulation can move at the speed of AI deployment.
OpenAI GPT-5.3 Details
Cerebras expansion scope and general availability timeline
Meta Privacy Backlash
Advocacy groups expected to respond to Name Tag
xAI Reorganization
Musk expected to announce structural changes
Anthropic Deployment Plan
$30B allocation strategy expected
EU AI Act Phase
New enforcement compliance deadlines
Final Thought
Valentine's Day 2026 is an appropriate moment to assess which AI companies are building relationships with the future and which are just going through the motions. The Great AI Divergence is not a temporary market correction. It is a structural separation between organizations with the clarity, talent, and capital to build what comes next and organizations that are increasingly defined by what they are losing.
The semiconductor industry approaching $1 trillion is the physical manifestation of this divergence. That trillion dollars of silicon is being purchased by the companies in Category One. The companies in Category Two are watching it happen and writing memos about timing.
Next week will test whether the divergence accelerates or whether any of the Category Two companies can change trajectory. History suggests that organizational inertia is harder to overcome than any technical challenge. But history also suggests that the AI industry does not follow historical patterns for very long.
Related Coverage
For deeper analysis of the funding dynamics driving this week's Anthropic story, see our analysis of Big Tech spending $650 billion on AI infrastructure.
Our prediction on AI startup talent exodus anticipated the xAI co-founder departures and tracks the broader pattern across the industry.
For context on how Anthropic's product expansion impacts enterprise software markets, see our coverage of Claude Cowork triggering a $285 billion SaaS selloff.