Good morning. The numbers are getting hard to parse today: Anthropic’s revenue run rate has apparently sextupled since December, Nvidia is writing $105 billion in credit for a single Ohio data center, and Qwen 3.8 27B is beating a model that was SOTA six months ago while running on a gaming PC. Meanwhile OpenAI just slashed GPT-5.6 Sol prices in half, which tells you something about the pricing pressure everyone is feeling.
Anthropic hits $65B annualized, IPO reportedly imminent. TechCrunch reports Anthropic’s annualized revenue crossed $65 billion at the end of July, up from $47B in May and $9B at the end of last year. Investors are modeling $100–120B for full-year 2026, and the company is preparing a possible fall IPO at a $2T+ valuation, which would be the largest market debut on record. For context, OpenAI recently reported $40B annualized — so the Claude-vs-GPT commercial gap has flipped and widened faster than most people expected.
Qwen 3.8 27B scores 52, beats Opus 4.6. Alibaba’s new 27B model ranks #1 in its size class on Artificial Analysis and matches DeepSeek V4 Flash, a >150B model. As one HN commenter put it: “It’s both funny and a bit terrifying” that a laptop-runnable model now beats what was broadly SOTA six months ago. Simon Willison’s hands-on confirms it’s excellent but flags the default “xhigh” reasoning setting as impractical — his SVG test burned 22,276 reasoning tokens over 21 minutes, versus two minutes with reasoning off for comparable output. Community workarounds mostly involve forked llama.cpp builds with manual reasoning flags.
GPT-5.6 Sol drops 50% in price. OpenAI cut Sol’s pricing on OpenRouter to $2.50/$15 per million tokens, positioning it against cheaper competitors like Grok 4.6. HN reaction is split — some love it, others insist every OpenAI release since 5.3-Codex has been a regression. One commenter noted that if OpenAI can cut Sol 50% and Luna 80% and stay profitable, the original margins must have been enormous. Related: Roboflow claims Sol is OpenAI’s best vision model ever, though the article buries that Gemini 3.5 Flash still beats it on most detection benchmarks while being cheaper and faster.
Groq pivots from chips to neocloud at a lower valuation. Groq raised $350M led by Disruptive with planned Nvidia participation, at a $3.5B valuation — half its $6.9B mark before Nvidia poached its founder and top talent in that $20B licensing deal. The company now runs 13 data centers serving 6M developers, and plans to scale from 54MW to 200MW+ by 2027, all on Nvidia hardware it once aimed to replace. The neocloud pivot mirrors CoreWeave’s, whose free cash flow struggles are the obvious cautionary tale.
Nvidia puts $1.5B into SB Energy for OpenAI’s Ohio site. TechCrunch reports Nvidia is investing $1.5B in SoftBank-linked SB Energy to lock in sole compute supply for OpenAI’s Ports-Pike facility near Cincinnati, plus up to $105B in credit for the project. The 8-gigawatt build will be powered by a $33B, 9.2-gigawatt natural gas plant on a former DOE uranium enrichment site. Natural gas plant costs are up 66% in two years, and analysts warn AI demand could triple gas prices in some regions.
Copilot Autofix opens a hole in Snowflake’s CI. Wiz Research disclosed that a Copilot-Autofix commit replaced a sanitized env
jq
pattern in Snowflake’s GitHub Actions workflow with direct shell string expansion, letting any GitHub user execute arbitrary commands and reach Snowflake’s internal Jira. The bug was five days old when Wiz’s autonomous agent found it; Snowflake patched same-day. The HN discussion settled on a structural point worth repeating: AI has cut the cost of introducing code changes dramatically, but the cost of reviewing them hasn’t budged. The bottleneck is moving.
Anthropic’s watermarks, and Gruber’s misfire. Anthropic detailed its adaptation of Google’s SynthID-Text, which biases token selection at low-stakes decision points to embed a detectable fingerprint — mostly to comply with the EU AI Act. John Gruber called it a “perversion of writing”, arguing LLMs should pick the “best” word every time. HN commenters noted this misses how sampling works — there is no single best token at temperature > 0, and the Gumbel softmax trick used here is provably distribution-preserving. A more interesting concern from the thread: verifying watermarks requires shipping your text to every provider that offers one.
That’s the day. The Anthropic revenue number is the one that will keep echoing — but if you’re looking for the more useful signal, it’s that a 27B open model just matched what was frontier six months ago, and OpenAI is cutting prices in half.