Weekly Download #38

Hi !

This week in Silicon Valley, startups & tech:

• Anthropic’s COBOL Kill Shot: A single blog post about Claude Code automating COBOL modernization wiped 13% off IBM ($30B+ gone in hours)

• Anthropic’s Dev Revolt: Sonnet 4.6 earns praise, but backlash over closed-source policies, no OpenClaw Oauth, and competitor bans is boiling

• OpenAI’s $850B Reality Check: A $100B+ round with Nvidia at $30B while Stargate stalls and compute targets shrink

• Distillation War Goes Hot: Anthropic accuses Chinese labs (DeepSeek, MiniMax, and Moonshot) of 16M+ ToS violations via API distillation

• The SaaS Extinction Trade: Software stocks are repricing in real time as AI disruption moves from theory to tickers

As we continue to build at Context Ventures, we've launched Startup Intros with a simple mission: to help early-stage founders find the right investors, faster and smarter.

📅 Startup Intros Events Coming Up

Feb 26: Agentic AI Founders' Night-Out: Pizza, fireside chat & networking with top AI commerce founders in SF. RSVP on Luma!

Feb 27: Caffeine & Capital @ Corgi Cafe: Casual Friday morning coffee & networking with 30-40+ founders & investors. RSVP on Luma!

🤖 Anthropic’s Most Contradictory Week in AI History

In one week, Anthropic shipped a top‑tier coding model, wiped $30B off IBM's market cap with a blog post, accused Chinese labs of distillation theft, got summoned to the Pentagon, and triggered the loudest dev revolt in AI history.

🧠 Claude Sonnet 4.6: The Model Is Genuinely Great

Anthropic released Claude Sonnet 4.6, and the benchmarks are strong. Artificial Analysis shows a massive intelligence bump, putting it right alongside Opus 4.6. Coding performance improved significantly, with broad upgrades across computer use, long-context reasoning, and design tasks.

The Pricing Wrinkle: Sonnet has historically been the affordable middle-tier model between Haiku (cheapest) and Opus (most expensive). But Sonnet 4.6 uses over 25% more reasoning tokens than Opus for the same tasks, which means the actual cost gap between Sonnet and Opus narrows to roughly 25% despite Opus costing 2x per token.

💀 The IBM Kill Shot: When a Blog Post Is a Weapon

Anthropic published a blog post outlining how Claude Code can automate COBOL modernization, which is the work IBM's consulting army charges premium rates for. Subsequently, IBM shares closed down 13.15%, erasing over $30B in market cap in a single session.

Why This Matters: IBM's mainframe and consulting businesses depend on the complexity and scarcity of COBOL expertise. There are estimated to be 220B lines of COBOL still running in production, powering banks, insurance companies, and government systems. Since many who understand it are retiring, IBM built an entire services business around this scarcity, but Claude Code is a huge threat.

Translation: In AI, a blog post can now move a $200B company, as this was a demonstration of Claude Code’s capability. It moved markets more than most acquisitions.

🔬 The Distillation War: 16M API Calls From China

Anthropic revealed that dominance has a target on its back. Anthropic says DeepSeek, MiniMax, and Moonshot violated its Terms of Service by prompting Claude a combined 16M+ times and using distillation to train their own products.

What Distillation Means: You query a frontier model millions of times, extract its reasoning patterns from the outputs, and use those patterns to train cheaper models that approximate the frontier model's capabilities. It's the AI equivalent of reverse-engineering a competitor's product by buying millions of units and taking them apart.

The Scale: 16M+ prompts aren't enthusiastic usage. It's industrial-scale extraction that is systematic, automated, and deliberate. This is the first time a major AI lab has publicly accused competitors of this practice.

The Geopolitical Irony: The US spent billions restricting chip exports to slow China's AI progress. If Chinese labs can just distill American frontier models remotely through API calls, hardware export controls become partially irrelevant.

The Enforcement Problem: How do you stop distillation when your product is literally an API? Rate limits help, but determined actors route through proxies and distributed accounts. This could accelerate the push for AI model IP protections for a legal framework the industry has barely begun to define.

🏛️ Pentagon Standoff Escalates

War Secretary Pete Hegseth summoned Dario Amodei to the Pentagon for what sources describe as a "likely tense" meeting over Claude usage. This follows last week's reporting that the Pentagon may sever its relationship with Anthropic over the company's refusal to build mass surveillance systems or autonomous weapons.

PAC Expansion: Meanwhile, Anthropic's political arm continues expanding: the Anthropic-backed super PAC Public First Action is running ads urging AI regulations in New Jersey, having raised nearly $50M with a target of $75M. Fighting the Pentagon while lobbying Congress is a bold two-front war.

😤 The Developer Revolt: "Anthropic Is a Cult"

Dev influencer Theo Browne published a detailed video this week, calling Anthropic "a cult" and documenting a pattern of antagonistic behavior toward the ecosystem that enabled Claude Code's dominance.

The Subscription Lockdown: Anthropic's $200/month Max plan subsidizes up to ~$2,700 in inference, a 15x subsidy. But Anthropic now explicitly prohibits using those subscriptions through any third-party tool, including its own Agent SDK, which its DevRel lead, Tar, publicly said was allowed in January. When devs asked for clarification, including Theo tagging Anthropic directly with a simple question about open-source apps, they got silence.

The Competitor Bans: Anthropic banned Windsurf (over an unconfirmed OpenAI acquisition rumor that didn't happen), banned OpenAI from benchmarking Claude, and banned xAI. If you compete, might compete, or use OpenClaw, you're cut off.

The OpenAI Contrast: In contrast, OpenAI employees engage publicly, contribute code to open-source competitors, and tell builders, "use our auth, build on our server." When OpenClaw needed the Codex integration, OpenAI devs helped make it work. Instead, Anthropic sent cease-and-desist letters, locked down policy, and served DMCA takedowns against anyone who shared source code that Anthropic had accidentally published.

The Influencer Problem: Multiple dev influencers report Anthropic pays roughly $3,000 for Claude Code ad content, then spends an estimated $1M+/day distributing those ads on Instagram by blasting creators' faces worldwide for a fraction of fair market value.

Translation: Anthropic built its dev moat on model quality, but moats require goodwill. The models are weapons-grade. The company treats devs like adversaries, competitors like enemies, and influencers like disposable assets. OpenAI isn't winning the model war, but it's winning the people war, and in a platform business, that might matter more.

📉 The SaaS Extinction Event Arrives

The AI-eats-software thesis stopped being theoretical this week. It showed up in stock tickers, earnings calls, and research notes. The repricing of the entire enterprise software sector is now happening in real time.

📊 The Broader SaaS Rout

Software stocks extended their weeks-long selloff as white-collar layoffs grow, the consumer economy withers, and social fabric frays. Atlassian's founders lost ~$7.2B in 2026, with TEAM down 45%+ YTD, the Nasdaq 100's worst performer.

Citrini's "2028 Global Intelligence Crisis" Thesis: The research firm published a scenario imagining an AI-driven crisis where white-collar layoffs grow, the human-centric consumer economy withers, and the social fabric frays. Dystopian? Maybe. But the market is pricing elements of it in right now.

⚖️ The Displacement Debate

Sam Altman acknowledged that there is both "real displacement by AI" and "AI washing," where companies blame AI for layoffs they would have made anyway. A CEPR study of 12,000+ EU companies found that AI adoption increases labor productivity by 4% on average, with no evidence of short-run job losses.

Software Development: Open-source projects like VLC and Blender are seeing a decline in the average quality of contributions, likely as AI coding tools lower barriers to entry. The tools that make code easier to write are also making code easier to write badly.

Translation: We've entered a phase where an AI company's blog post is a more effective weapon than a product launch. The SaaS repricing isn't about whether AI will disrupt enterprise software. It's about how fast, and the answer this week was "in a single trading session."

💰 OpenAI's $850B Reality Check

OpenAI is closing in on the largest private fundraising in history, but the math behind the hype is getting harder to defend.

🏦 The Math: $100B+ Round & $280B Revenue by 2030

OpenAI is close to finalizing its record $100B+ raise at a valuation that could exceed $850B. Nvidia is putting in up to $30B, as it’s massive, but a $70B haircut from the $100B commitment discussed last year that stalled after internal doubts at the chip giant.

• The Revenue Math: OpenAI projects $280B+ in revenue by 2030, up from $20B+ today. That's 14x growth in four years. Google took 15 years to hit $280B.

• The Quiet Haircut: OpenAI is now targeting ~$600B in total compute spend by 2030, months after Altman touted $1.4T in total compute spend. A 57% reduction that nobody's explaining.

• Enterprise Play: OpenAI launched Frontier Alliances with Accenture, BCG, Capgemini, and McKinsey, letting consulting firms sell AI transformation to the Fortune 500 and taking a cut of every deployment.

🚧 Stargate Stalls & Safety Overhang

The Execution Gaps: Stargate stalled amid clashes with SoftBank, and building its own data centers isn't a near-term priority. Instead, 200+ people are working on AI devices such as a $200-$300 smart speaker with a camera, possibly glasses, or a smart lamp. Altman called Musk's space data center idea "ridiculous."

The Safety Overhang: Staff flagged a Canadian mass shooting suspect months before the attack, but didn't report to police. Meanwhile, Benedict Evans argues OpenAI has a large user base but narrow engagement, and incumbents are catching up fast.

Reality Check: OpenAI is raising at $850B while cutting compute targets by half, stalling on Stargate, and pivoting to $200 smart speakers. The valuation assumes a future that even OpenAI's own numbers can't consistently describe.

⚡ Startup Quick Hits

🌐 World Labs' $1B at ~$5B: Fei-Fei Li's spatial AI startup raised from AMD, Nvidia, Autodesk ($200M), a16z, and Fidelity. When Autodesk invests and advises, spatial AI is commercial infrastructure.

💰 Vestwell's $385M Series E at $2B: Workplace savings platform doubled valuation with Blue Owl and Sixth Street. 2M+ savers, $50B in assets, $200M+ ARR, growing profitably.

⚙️ Temporal's $300M Series D at $5B: A16z led for the "durable execution" platform keeping AI agents from crashing mid-workflow. OpenAI uses it in production. Revenue up 380% YoY.

🔧 Taalas Raises $169M: Hardwires AI models directly into custom silicon for 73x faster inference than Nvidia's H200 at 1/10th the power. Quiet Capital and Fidelity backed.

💻 Code Metal's $125M Series B at $1.25B: AI-powered verified code translation for defense and enterprise. Salesforce Ventures led. IBM's stock drop just made this market a lot more interesting.

⚡ Ubicquia's $106M Series D: AI-driven grid monitoring for 1,000+ utilities and municipalities. 67 Capital and Marunouchi led. When every data center needs power, grid management becomes essential.

🛰️ Aalyria's $100M Series B at $1.3B: Google spinoff orchestrates traffic across satellite, air, and ground networks. Battery Ventures led. Space-based internet infrastructure keeps attracting capital.

🔬 FreeForm's $67M Series B: AI-native 3D printing using 18 lasers to fuse metal powders into precision components. Founders Fund and Nvidia's NVentures backed. Hardware manufacturing meets AI.

👷 Humand's $66M Series A: HR app for 1.6M+ deskless workers across 1,500 companies in 51 countries. Goodwater and Kaszek led. Enterprise software for people who don't sit at desks.

🍕 Pepper's $50M Series C: Digital storefront for 500+ independent food distributors handling $30B in annual GMV. Lead Edge Capital led, bringing total funding to $100M.

💰 Investor Quick Hits

🔍 Garry Tan's Anthropic Insight: ~50% of Anthropic's AI agent tool calls are software engineering. Every other vertical is under 5%. If you're building an AI agent outside of coding, you might have less competition than you think.

📊 SemiAnalysis' Dylan Patel Eyes VC Fund: Early talks for a fund potentially in the hundreds of millions, having already raised a $50M SPV toward Fluidstack's $700M round. Analyst-to-investor pipeline accelerating.

☁️ Google Eyes $100M Fluidstack Investment at $7.5B: Also backstopping crypto miners building data centers to expand TPU usage. When you can't build fast enough, fund others to build for you.

🇮🇳 Peak XV's $1.3B Post-Sequoia Independence Fund: First major raise since splitting from Sequoia in 2023, now managing $10B+ across 450+ portfolio companies. AI, fintech, and consumer-focused across India and APAC.

💸 IPO & M&A Quick Hits

🖥️ Cerebras Confidentially Files for US IPO: The AI chipmaker refiled after withdrawing last year. Listing could come as soon as Q2; just as the AI hardware market enters its most competitive phase.

🅿️ Uber Acquires SpotHero: Adds parking reservations across 13,000+ locations in 400+ cities. Uber's strategy: own every step between "I need to go somewhere" and arriving.

💳 PayPal Attracts Takeover Interest: Stock down ~85% from 2021 highs, market cap ~$38B. At least one large rival is eyeing the whole company. When a fintech pioneer with 434M accounts trades at multi-year lows, sharks circle.

⚛️ IQM Going Public via SPAC at $1.8B: Finnish quantum computing company merging with NJ-based Real Asset Acquisition Corp. 21 systems sold, $35M 2025 revenue. Quantum SPACs in 2026; some patterns never die.

📸 Lightricks Splits Consumer and GenAI Units: Facetune ($100M profit) separating from GenAI unit LTX ($150M invested, 3M downloads). Investors pay premiums for pure-play AI, not diversified software.

⚡ SoftBank Plans $33B Ohio Power Plant: 9.2 GW facility would be the largest US power plant, part of $550B US-Japan framework. When your AI strategy requires building your own power plant, that's nation-building.

🌟 Editor's Note

At Startup Intros, our mission is to bring the latest founder-investor news straight to your inbox, keeping you ahead in the fast-paced world of Silicon Valley.

💭 Parting Thoughts: This Week was about Boundaries

Anthropic shipped the best coding model on the market, cratered IBM with a blog post, accused China of industrial-scale model theft, and alienated the developer community that built Claude Code's dominance, all in one week. The models are weapons-grade. The developer relations are in freefall. OpenAI isn't winning the model war, but it's winning the people war.

Meanwhile, the entire SaaS sector is getting repriced based on blog posts that haven't even become products yet. The question isn't whether AI disruption is real; IBM's chart answered that Monday. The question is whether anyone has priced it correctly.

Forward to a friend or hit reply to let me know what you're seeing in your world.

Till next time!

Till next time!

P.S. Raised this week, and we missed you? Want to be featured? Have tips or funding questions? Reply or DM us as we’re here to help.

Startup Intros Weekly Download: Your trusted source for founder-investor insights, delivered with clarity and focus.