THE AI RACE

Anthropic catches OpenAI using Claude to build GPT-5

Anthropic pulled the plug on OpenAI's Claude access last week after discovering that OpenAI engineers were systematically using Claude Code to benchmark and improve GPT-5, violating terms of service just weeks before the model's expected August launch.

OpenAI had integrated Claude directly into internal development tools through API access, running systematic comparisons across coding tasks, creative writing and even sensitive safety scenarios involving content like self-harm and defamation. Anthropic's terms explicitly prohibit using their service to "build competing products" or "train competing AI models." 

OpenAI called the practice "industry standard" benchmarking. Spokesperson Hannah Wong said they were "disappointed" considering their API remains available to Anthropic. But Anthropic spokesperson Christopher Nulty wasn't buying it: "This is a direct violation of our terms of service."

Back in June, Anthropic cut off Windsurf's Claude access after OpenAI tried to acquire the coding startup for $3 billion. Microsoft blocked the deal over IP rights concerns, then Google hired Windsurf's CEO and key staff for $2.4 billion while Cognition acquired what remained of the company.

Corporate surveillance and data warfare have deep roots in tech:

• Facebook's "Project Ghostbusters" used Onavo VPN to spy on Snapchat, intercepting encrypted traffic from 33 million users to inform Instagram Stories development and acquisition decisions

• Facebook blocked Twitter's Vine app in 2013, with Mark Zuckerberg personally approving the decision within hours of launch

• Twitter's API purge decimated apps like Block Party, forcing complete business pivots when access disappeared overnight

• Countries are now blocking DeepSeek while hundreds of companies ban the Chinese AI over espionage concerns

There's something deeply satisfying about watching OpenAI get caught red-handed like this. Using your competitor's best coding tool to build your own competing product, then calling it "industry standard" when you get cut off, is peak Silicon Valley behavior.

These companies are spending billions on compute and talent to build these systems. Why would you hand your competitors a roadmap to beat you? The collaborative early days of AI development are over. We're now in the era where models are treated like nuclear secrets.

TOGETHER WITH PENDULUM

How To Boost Your GLP-1 Naturally

GLP-1 is all the rage these days, and for good reason – it’s the hormone that tells your brain that you’re full, which helps manage appetite and cravings. But what if we told you there was a way to boost your GLP-1 entirely naturally?*

Enter Pendulum and their Akkermansia supplement (which Deep View readers can save 20% on right here).

Akkermansia is a keystone probiotic in your body that can reduce by up to 90% as you age, and Pendulum’s team of top scientists – from schools like Johns Hopkins, Harvard, Stanford, and Berkeley – have manufactured the first-ever live strain of it. With 100 million active units per dose, it can do wonders for your gut lining, metabolic health, and natural GLP-1 production.*

Want to try Pendulum for yourself? Get 20% off your Akkermansia probiotic right here – and yes, it’s HSA/FSA eligible!

*Based on preclinical studies. Not intended for weight loss. 

ECONOMY

AI infrastructure spending rivals 19th-century railroads

AI capital expenditure now represents roughly 2% of US GDP, according to analyst estimates based on Nvidia's datacenter revenue. That's around $520 billion annually, putting AI infrastructure spending within shouting distance of peak 19th-century railroad construction.

The scale becomes clearer when you consider that without AI datacenter investment, Q1 GDP contraction could have been closer to -2.1% instead of the mild downturn we actually saw. We're essentially running a massive private sector stimulus program disguised as an AI buildout.

This explains one of the more puzzling aspects of the current economy. Despite tariffs, political uncertainty, and the usual economic headwinds, growth has held up better than expected. The answer appears to be that Big Tech is spending so aggressively on GPUs and datacenters that it's masking underlying weakness.

The scale becomes clear when you look at individual company commitments. Microsoft plans $80 billion for AI datacenters in fiscal 2025, Amazon expects over $100 billion, Google allocated $75 billion and Meta budgeted $60-65 billion. Combined, these four companies alone are spending over $315 billion this year.

But that money is coming from somewhere else. Non-life science venture capitalists are primarily funding AI startups at present. Amazon recently announced cloud layoffs driven by spiraling AI costs. Manufacturing and traditional infrastructure are being starved for capital as it gets rerouted to data centers.

• Even Xi Jinping is sounding the alarm, with 250+ datacenters under construction in China and the Chinese president threatening intervention.

• The spending has already surpassed peak telecom capex during the dot-com bubble as a percentage of GDP.

The uncomfortable truth is that we're building short-lived infrastructure on rapidly depreciating technology curves, not century-spanning railroads. And when this particular spending spree ends, the economic support it's providing disappears with it.

TOGETHER WITH DELVE

Get Compliant in 15 Hours—Not 6 Months

Delve helps fast-growing companies get compliant with an AI-automated platform. SOC 2, HIPAA, and any other certification in just 15 hours.

The proof is in the pudding:

• Lovable → From zero to fully SOC 2 compliant in just days

• Cluely → Unlocked millions in enterprise contracts within just days

• Bland → Switched over, got SOC 2 compliant and unlocked $500k ARR within 7 days

If you’re suffering elsewhere, they’ll even migrate you off another platform all included

Deepview readers get $1,000 credit using code DEEPVIEW1KOFF + Free Airpods Pros when you get compliant

Getting compliant in days >>> quarters. Book a 30-minute demo to get started

KEY PLAYERS

Researcher who almost joined OpenAI for $175K turns down $1.5 billion from Meta

Andrew Tulloch, a co-founder at Thinking Machines Lab (TML), just rejected a billion-dollar package from Mark Zuckerberg that could have reached $1.5 billion over six years with bonuses and stock performance.

This is the same researcher who almost joined OpenAI back in 2016, when Greg Brockman wrote to Elon Musk that Tulloch was "very close to saying yes" but "concerned about taking such a large paycut" from his $800,000 Facebook salary to OpenAI's $175,000 base with $125,000 bonus.

• Eight years later, Tulloch finally made the jump to work for Mira Murati at TML, where he's now wealthy enough to turn down generational wealth from his former employer

• Not a single TML researcher took Zuckerberg's offers, despite Meta approaching more than a dozen employees at the roughly 50-person startup

• Tulloch's background makes this particularly brutal for Meta. The Australian researcher graduated with the highest GPA among science students at the University of Sydney, became a distinguished engineer at Facebook, and was considered "an extreme genius" by former Facebook executive Mike Vernal

When Murati left OpenAI in September, more than 20 colleagues followed her to TML, including co-founder John Schulman. The loyalty runs deeper than just Murati's leadership style.

TML recently raised funding at a $12 billion valuation, meaning researchers hold equity that could appreciate far beyond Meta's cash offers. When your startup is worth more than many Fortune 500 companies, walking away from even a billion-dollar corporate job becomes much easier.

We've moved from an era where Big Tech could outbid anyone to one where the most valuable researchers control billion-dollar startups themselves. Zuckerberg is no longer just competing against other companies for talent — he's competing against the wealth his targets have already accumulated. And apparently, that's a fight even a billion dollars can't win.

LINKS

• Why open-source AI became an American national priority

• Apple’s new ‘Answers’ team eyes ChatGPT-like product in AI push

• The uproar over Vogue’s AI-generated ad isn’t just about fashion

• Lina Khan points to Figma IPO as vindication of M&A scrutiny

• Fundamental Research Labs nabs $30M+ to build AI agents across verticals

• Google releases Gemini 2.5 Deep Think for AI Ultra subscribers

• Senate confirms Sean Cairncross to be national cyber director under Trump

• Amazon is toying around with putting ads in Alexa+

• Apple has now shipped 3 billion iPhones

• Gemini 2.5 Deep Think: Version of the model that achieved an IMO gold medal, using parallel "thinking time"

• Ollama Desktop App: Ollama’s macOS and Windows now include a way to download and chat with models

• Mixio: Create an AI avatar of yourself that streams 24/7 (this one feels a bit Black Mirror?)

• Delve: AI agents that help you get compliant with SOC 2, HIPAA and more overnight*

• Synthesia: Lead Brand Designer

• Meta: Product Manager, Business AI

• People.ai: Staff Product Manager, SalesAI Forecasting

• Writer: Product manager, agent builder

GAMES

A QUICK POLL BEFORE YOU GO