AI Daily Digest – March 6, 2026

Good morning, the government tried to kill Anthropic and accidentally made it the most popular AI app in the world, OpenAI dropped its most powerful model yet, SoftBank is borrowing $40 billion just to double down on its OpenAI bet, and Broadcom just told Wall Street it expects $100 billion in AI chip revenue by next year. Here’s what happened 👇

1. The Pentagon Labeled Anthropic a Security Risk. It Backfired Spectacularly.

On Thursday, the US Department of Defense officially designated Anthropic a “supply-chain risk” — a formal government label that has caused defense contractors to preemptively drop Claude “out of an abundance of caution.” Palantir, one of the Pentagon’s closest AI partners, is now scrambling to rip Anthropic out of its own military software. The designation limits Claude’s use specifically on contracts directly with the Department of War, though Anthropic says the vast majority of its customers are unaffected.

But here’s the twist that nobody in Washington planned for: Claude has been breaking daily signup records in every country where it’s available since early last week — and as of this morning, it’s topping the App Store charts for free apps and AI apps across dozens of countries, including the US, Canada, and most of Europe. The designation meant to sideline Anthropic turned into its best marketing campaign in company history.

CEO Dario Amodei confirmed in a public blog post that Anthropic will challenge the Pentagon’s designation in court. He said the language in the DoD’s letter “plainly applies only to the use of Claude by customers as a direct part of contracts with the Department of War, not all use of Claude by customers who have such contracts” — meaning the ban is narrower than the headlines made it sound. Palantir, one of the Pentagon’s closest AI partners, is nonetheless scrambling to remove Anthropic from its military software stack.

Why it matters: This story has moved from a policy dispute into something more fundamental — a public referendum on whether AI companies should have ethics guardrails, and whether the government can punish them for it. The fact that regular people responded by downloading Claude in record numbers suggests the answer, at least in the court of public opinion, is yes.

Sources: The Verge | TechCrunch | Reuters

2. OpenAI Drops GPT-5.4 — Its Most Capable Model for Professional Work

While the Anthropic drama dominated headlines, OpenAI quietly released its most capable model yet on Thursday. GPT-5.4 comes in three flavors: a standard version, a reasoning-focused “Thinking” version, and a performance-optimized “Pro” version. OpenAI is billing it as “our most capable and efficient frontier model for professional work.”

The numbers are impressive. GPT-5.4 scored 83% on OpenAI’s own GDPval benchmark for knowledge work tasks — things like financial modeling, legal analysis, and slide deck creation. It’s 33% less likely to make factual errors in individual claims compared to GPT-5.2. The API version supports a context window of 1 million tokens, by far the largest OpenAI has offered — meaning it can hold an entire novel, a full codebase, or months of meeting transcripts in a single conversation. It also set new records on computer use benchmarks OSWorld-Verified and WebArena, which test AI agents’ ability to operate computers directly.

For developers building AI applications, GPT-5.4 introduces “Tool Search” — a new system where the model looks up tool definitions only when needed, instead of loading all tools upfront. In systems with hundreds of available tools, this cuts both cost and latency significantly.

OpenAI also addressed one of AI safety’s biggest open questions: whether reasoning models misrepresent their “chain of thought” — the step-by-step thinking visible during complex tasks. Testing on the Thinking version shows lower rates of deceptive reasoning, with OpenAI claiming the model “lacks the ability to hide its reasoning.”

Why it matters: GPT-5.4 is arriving at a moment when OpenAI badly needs to remind people why they came to it in the first place. The 1M token context window and agent benchmarks hint at what’s next: AI that can work on a problem for hours, not seconds, handling the full scope of a complex professional task in one session.

Sources: TechCrunch | The Verge

3. SoftBank Is Borrowing $40 Billion Just to Invest More in OpenAI

This one arrived this morning and the number alone demands explanation: Japanese conglomerate SoftBank is seeking a bridge loan of up to $40 billion — primarily to finance its investment in OpenAI, Bloomberg News reported Friday. JPMorgan is among four banks underwriting the facility. The loan would have a roughly 12-month tenor, meaning SoftBank plans to repay it within a year, presumably after OpenAI goes public or after other funding events materialize.

To understand why this number is staggering: SoftBank already holds about 11% of OpenAI. Last month, it put in $30 billion as part of OpenAI’s $110 billion funding round — a round that also included $50 billion from Amazon and $30 billion from Nvidia, and valued OpenAI at $840 billion. OpenAI is simultaneously laying the groundwork for an IPO that could push its valuation toward $1 trillion. CEO Masayoshi Son has publicly described his OpenAI position as going “all in.”

To put the $40 billion in perspective: it is roughly equal to the entire GDP of Honduras. It’s more than Google paid for all acquisitions combined in 2024. SoftBank is borrowing an amount larger than most countries’ annual budgets to increase a bet on a single AI company that didn’t exist 10 years ago.

Why it matters: The AI investment cycle isn’t slowing down — it’s accelerating into territory that requires entirely new vocabulary. At some point the math has to close: OpenAI hit $25 billion in annualized revenue as of last month, up from nearly zero two years ago. But at a $1 trillion valuation, the implied multiple is extraordinary. SoftBank is betting the trajectory holds. The world is watching whether it does.

Sources: Reuters

4. Trump May Force Every Country to Invest in US Data Centers to Buy AI Chips

Reuters obtained a draft document from the Trump administration outlining a sweeping new framework for AI chip exports — and it’s a major departure from everything before it. The core idea: if you want to buy more than 200,000 advanced AI chips from US companies like Nvidia or AMD, your government may need to invest in US AI data centers first. Even small purchases under 1,000 chips could require a license. Orders of up to 100,000 chips would require government-to-government security assurances.

This flips the Biden-era approach on its head. Biden’s “AI diffusion rules” exempted close US allies — countries like the UK, Japan, and South Korea — from most chip export restrictions. Trump is treating everyone the same: ally or not, if you want chips, you negotiate with Washington first. The framework already exists in practice: Saudi Arabia and the UAE both agreed to invest in US AI infrastructure in exchange for chip access. Trump is now looking to formalize that as the global standard.

The draft also notably does not restrict exports of AI model weights — the core parameters of a trained AI system — which Biden had moved to protect. That omission could allow foreign entities to more freely access the underlying intelligence of advanced AI models, not just the hardware.

“The rule could help address chip diversion to China,” said Saif Khan, a former Biden national security official, “but the license requirements are overly broad — raising concerns the administration intends to use the controls as negotiation leverage with allies rather than for security.”

Why it matters: The US currently has something close to a monopoly on the most advanced AI chips, and this proposal would turn that monopoly into explicit geopolitical leverage. Want to build AI infrastructure in your country? First, invest in America. The global AI race just became inseparable from global trade and foreign policy. Every country with AI ambitions — Europe, India, Japan, South Korea — now has to weigh chip access against sovereignty.

Sources: Reuters

5. Broadcom Just Told Wall Street It Expects $100 Billion in AI Chip Revenue by 2027

While Nvidia dominates the headlines, Broadcom quietly dropped one of the most bullish earnings reports in the AI hardware space this week. Q1 AI revenue came in at $8.4 billion — more than double the same period last year. Total revenue rose 29% to $19.31 billion. And then CEO Hock Tan said something that stopped analysts mid-sentence: “Today, in fact, we have line of sight to achieve AI revenue from chips in excess of $100 billion in 2027.”

To understand why this matters, you need to understand what Broadcom actually does. It doesn’t sell AI chips off a shelf like Nvidia. Instead, it works with Big Tech companies to design their custom AI processors — the chips Google calls TPUs, the custom accelerators Meta and OpenAI are building in-house. Broadcom does the hard engineering work of turning an early design into a manufacturable chip, then TSMC fabrics it. The clients pay Broadcom for the design work and buy the chips at scale.

This week’s numbers revealed the scale of those relationships. Broadcom is delivering 1 gigawatt’s worth of custom AI chips to Anthropic in 2026 alone — rising to 3 gigawatts in 2027. It will ship OpenAI’s first custom processor in 2027 as well. AMD separately disclosed deals approaching 6 gigawatts with Meta and OpenAI. Nvidia disclosed 5 gigawatts to OpenAI last week. The unit of measurement for AI infrastructure is now gigawatts — the same unit used for power plants.

Marvell Technology, another chip designer focused on AI data center interconnects, also reported this week and forecast multi-year AI chip growth. Its shares jumped 15%.

Why it matters: The AI chip story is no longer just “Nvidia vs. everyone.” Broadcom, AMD, and Marvell are all posting massive numbers, all forecasting growth for years out, and all building custom silicon for the same handful of hyperscalers. The AI hardware market is expanding fast enough for multiple $100B players to coexist — and the investment required to build it is measured in the same units as the electrical grid.

Sources: Reuters | Reuters — Marvell

Quick Hits

• Oracle is cutting thousands of jobs despite being OpenAI’s biggest cloud partner: Oracle has a $30 billion/year cloud deal with OpenAI — but the cost of building the data centers needed to support it is straining the company’s finances, Bloomberg reported. Oracle is planning “thousands” of job cuts as it tries to manage a cash crunch. The AI infrastructure buildout is minting winners and victims at the same time, sometimes in the same company. (Reuters)

• Netflix bought Ben Affleck’s AI filmmaking startup: Netflix acquired InterPositive, a company Affleck co-founded to build AI-powered tools for movie production. Affleck is joining Netflix as a senior adviser. AI is arriving in Hollywood not as a replacement for filmmakers — but as a tool being built and sold by them. (Reuters)

• Meta’s AI glasses were sending intimate footage to human reviewers in Kenya: CNBC and The Verge reported that footage captured by Ray-Ban Meta smart glasses — including sensitive and sometimes intimate content — was reviewed by human contractors in Kenya. Meta is now facing a lawsuit over the privacy implications. Meta separately agreed to temporarily allow competing AI chatbots on WhatsApp in the EU to stave off antitrust action. (The Verge)

• A new open-source AI was trained on trillions of DNA base pairs: Researchers published a large genome model capable of identifying genes, regulatory sequences, splice sites, and more — trained on a scale that wasn’t possible a few years ago. It’s the biology equivalent of a foundation model. The implications for drug discovery and genetic medicine are significant. (Ars Technica)

• UK House of Lords says AI companies must license creative work before training on it: A UK parliamentary committee recommended a “licensing-first” approach to AI training data — meaning AI labs would need permission before scraping books, music, and articles, rather than treating it as a fair-use free-for-all. This directly conflicts with how most major AI models were built. (Reuters)

That’s it for today. This week’s AI story has two distinct threads running in opposite directions: the technology keeps getting more powerful (GPT-5.4, $100B chip forecasts, $40B bets on a single company), while trust in the institutions building it keeps eroding (Pentagon battles, leaked memos, glasses that spy on you). At some point those threads have to cross. This week, they’re still pulling apart.

Forward this to someone who needs to stay in the loop.