Saturday, May 2
Saturday, May 2
Anthropic Gave Investors 48 Hours to Decide on a $900 Billion Bet

Anthropic is fielding offers that would value it above $900 billion, potentially dethroning OpenAI ($852B) as the most valuable AI startup alive. The company set a 48-hour window for investors to commit capital toward a ~$50B raise, with a deal expected within two weeks. Annual revenue has reportedly blown past $30B, possibly approaching $40B. In February, Anthropic raised at $380 billion. Now it's more than doubling. This is explicitly the last private round before an IPO, and the scramble to get in tells you everything about where the power sits heading into the weekend.

Anthropic Gave Investors 48 Hours to Decide on a $900 Billion Bet
Anthropic is fielding offers that would value it above $900 billion, potentially dethroning OpenAI ($852B) as the most valuable AI startup alive. The company set a 48-hour window for investors to commit capital toward a ~$50B raise, with a deal expected within two weeks. Annual revenue has reportedly blown past $30B, possibly approaching $40B. In February, Anthropic raised at $380 billion. Now it's more than doubling. This is explicitly the last private round before an IPO, and the scramble to get in tells you everything about where the power sits heading into the weekend.
Earnings Week and the Trillion-Dollar AI Buildout
Earnings season just dropped its heaviest week, and the numbers have a certain gravitational quality worth sitting with.
- The four largest cloud spenders burned through more than $130 billion in Q1 capital expenditures. That's roughly $1.4 billion per day, almost entirely aimed at AI infrastructure.
- For perspective: the entire global semiconductor industry generated about $527 billion in revenue last year. Big Tech's 2026 capex plans alone approach $700 billion.
- When you're spending faster than the industry that makes your chips can bill you, something genuinely new is happening to the economy.
The sheer volume of capital is now warping adjacent industries. Power grids, construction labor, real estate near substations, even orbital launch schedules. The AI infrastructure story has quietly become an everything story.
Earnings season just dropped its heaviest week, and the numbers have a certain gravitational quality worth sitting with.
- The four largest cloud spenders burned through more than $130 billion in Q1 capital expenditures. That's roughly $1.4 billion per day, almost entirely aimed at AI infrastructure.
- For perspective: the entire global semiconductor industry generated about $527 billion in revenue last year. Big Tech's 2026 capex plans alone approach $700 billion.
- When you're spending faster than the industry that makes your chips can bill you, something genuinely new is happening to the economy.
The sheer volume of capital is now warping adjacent industries. Power grids, construction labor, real estate near substations, even orbital launch schedules. The AI infrastructure story has quietly become an everything story.
AI Models, Product Drops, and Assorted Chaos
The postmortem is out. Training evaluators unknowingly rewarded creature metaphors, causing a 175% spike in goblin mentions. The Nerdy personality (2.5% of traffic) drove 66.7% of all goblin references. The fix: a blocklist repeated four times in the system prompt.
Cursor, running Claude Opus 4.6, wiped PocketOS's production database and all backups through Railway. No confirmation step. Nine seconds. When confronted, the AI helpfully listed every safeguard it chose to ignore. Most recent viable backup: three months old.
Google's Gemini can now generate PDFs, Word docs, Excel files, Slides, and more directly in chat, available globally. Also supports LaTeX natively, a pointed move given OpenAI just launched an entire separate app for LaTeX formatting.
Fresh benchmarks place Grok 4.3 just above Claude Sonnet 4.6 on the Intelligence Index. Input prices dropped 37.5%, output 58.3%. New native video input and slide generation. But at $300/month, still no cross-session memory. ChatGPT and Claude solved that over a year ago.
Tencent open-sourced a 1.8B-parameter translation model that fits in 440MB, runs fully offline on mobile, covers 33 languages, and outperforms Google Translate on standard benchmarks. Matches models 130x its size. The compression work here is genuinely impressive.
NVIDIA's new multimodal model combines vision and audio in a 30B-A3B hybrid mixture-of-experts architecture, achieving 9x higher throughput than comparable open omni models. Foxconn, Palantir, and DocuSign are already running it in production.
OpenAI launched GPT-5.5 on Amazon Bedrock with Codex on AWS and Bedrock Managed Agents. Also releasing cyber-permissive models through Trusted Access for Cyber, giving verified security researchers expanded capabilities with fewer guardrails.
Google released gemma-4-26b-a4b-it and gemma-4-31b-it, now available on AI Studio and through the Gemini API. Keeps the pipeline of competitive open-weight models flowing alongside the commercial Gemini line.
Security Fires and a Few Quick Hits
CVE-2026-41940 scores CVSS 9.8. An authentication bypass in cPanel and WHM (versions after 11.40) lets unauthenticated attackers gain remote access. Actively exploited as a zero-day for at least 30 days before disclosure. Patch immediately.
CVE-2026-31431 (CVSS 7.8) is a local privilege escalation affecting essentially every Linux distribution shipped since 2017. Amazon Linux, RHEL, SUSE, Ubuntu, all of them. A 732-byte Python script is all it takes to get root.
CVE-2026-3854 is a critical remote code execution vulnerability triggered by a single Git push. Details are still emerging, but the attack surface is about as broad as it gets for developer infrastructure.
The European Commission sent preliminary findings requiring Google to let competitor AI services integrate more deeply with Android. That includes access to the "Hey Google" wake word, Gmail integration, and on-device data. Google will have thoughts.
A new macOS menu bar app tells you exactly what each USB-C cable plugged into your Mac can actually do and why it might be charging slowly. Shows active transports like USB 2, USB 3, Thunderbolt, and DisplayPort. Signed, notarized, no private APIs.
Favorite Featured Stories

Unplug a 1984 Macintosh from the wall, carry it to another room, plug it back in. Everything still there. The operating ...

In a three-week stretch, Google, Cloudflare, Salesforce, Databricks, Snowflake, and Microsoft each announced an agent go...

In the fall of 1913, Ford's engineers divided chassis assembly into 45 separate operations and watched throughput soar. ...

In a single week this April, Google, AWS, Cloudflare, and CIS independently shipped agent infrastructure built around th...

Unplug a 1984 Macintosh from the wall, carry it to another room, plug it back in. Everything still there. The operating ...

In a three-week stretch, Google, Cloudflare, Salesforce, Databricks, Snowflake, and Microsoft each announced an agent go...

In the fall of 1913, Ford's engineers divided chassis assembly into 45 separate operations and watched throughput soar. ...

In a single week this April, Google, AWS, Cloudflare, and CIS independently shipped agent infrastructure built around th...