The numbers

Etched’s valuation has roughly quadrupled in eight months: $5 billion in December 2025, $10.3 billion after a $300 million Series C in July 2026, and now $21 billion with this latest $700 million round, led by Jane Street. Other backers named in TechCrunch’s reporting include Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, and Blackstone.

What Etched actually builds

Etched makes AI inference hardware sold as complete systems it calls “frontier inference clusters,” built around two custom components: a prefill chip that processes prompts at low voltage, and cluster-scale memory that lets multiple chips share memory at high speed with low latency. Co-founder and COO Robert Wachen described the split to TechCrunch: “Inference is built in two stages, prefill and decode,” with prefill handling prompt comprehension and decode generating the output tokens.

Jane Street, a trading firm rather than a traditional venture fund, framed its investment around direct testing: “We tested the chip and are pleased with the early results. Etched’s unique approach to inference delivers the precision we will need to support our most demanding workloads.”

Why it matters

A trading firm testing a chip and then leading its funding round, rather than a fund built around VC-style bets, stands out here: it suggests Etched’s pitch is landing with a buyer that would actually run its own workloads on the hardware, not just underwrite the story. Quadrupling a valuation in eight months is an extraordinary pace even by 2026’s AI-infrastructure standards, and it’s worth watching whether that reflects real deployed demand for Etched’s inference clusters or the broader capital-availability dynamic driving valuations across the sector, since TechCrunch’s reporting covers the round itself, not independent customer-deployment numbers.

What to watch: whether Etched discloses actual production customers, not just investor validation, in the months following this round.

More Startups & Funding coverage, or everything tagged funding.

What Alibaba actually shipped

Qwen3.8-27B is a dense, 27-billion-parameter model that Alibaba says matches the performance of Qwen3.7-plus, a mixture-of-experts model roughly ten times its size, according to Alibaba’s own release post. Alongside it, Alibaba released Qwen3.8-2.4T-A95B, a 2.4 trillion-parameter MoE model that activates 95 billion parameters at a time, the flagship of the family. Both carry a native 262K token context window that extends to 1 million, native vision-language support for images and video, and Apache 2.0 licensing. Weights are published on Hugging Face and ModelScope.

How it’s landing

Within two days of release, Qwen3.8-27B became one of the top five most-liked models on Hugging Face, per Alibaba’s own count. The flagship model placed third globally on Arena AI’s CodeArena for front-end web development and third on Artificial Analysis’s Agentic Index, third-party benchmark placements Alibaba cited in its release post rather than figures independently verified here.

Why it matters

CNBC framed the release explicitly as Alibaba answering Meta’s open-model push, and the laptop-ready framing looks like a direct pitch to developers who don’t want a network dependency, or a hosted-API bill, in their coding loop. Alibaba’s cited growth numbers, more than 460 open-sourced models, over 300,000 derivative models, and 3 billion-plus downloads, suggest the open-weight strategy is compounding rather than a one-off release, though those figures come from Alibaba’s own count, not an independent audit. Whether the “runs on a laptop” claim holds up outside Alibaba’s own benchmarks is the detail worth watching next: independent testing will settle that faster than any vendor’s blog post.

What to watch: whether Meta or another open-weight lab answers with its own laptop-class release, and whether benchmarks run outside Alibaba’s own citations confirm the local-hardware performance claims.

More AI News coverage, or everything tagged qwen.

The trajectory

Per CNBC’s reporting, Anthropic’s run rate moved from roughly $9 billion at the end of 2025 to more than $30 billion in April, $47 billion in May, and $65 billion by the end of July. The company also shared a preliminary second-quarter revenue figure of $11.5 billion with investors, according to the same reporting.

How it compares to OpenAI

OpenAI’s most recently disclosed run rate is $40 billion, from an internal message by co-founder Greg Brockman that Axios reported separately. The two companies may calculate “run rate” differently, which complicates a direct comparison, but the gap is large enough that several outlets framed it as Anthropic overtaking OpenAI on this specific metric.

The IPO context

Anthropic is reportedly meeting with potential new investors ahead of a planned IPO expected in September or October, working with Morgan Stanley, Goldman Sachs, and JPMorgan on the offering. The company reached a $965 billion valuation after a funding round in May; the new revenue figures give it a fresh data point to defend that number to prospective public-market investors.

Why it matters

None of this comes from an Anthropic press release or a regulatory filing. It’s sourced to what the company told investors privately and what reporters subsequently obtained, so the exact figure should be read as investor-communicated, not audited. That caveat aside, the trajectory itself, not just the single $65 billion number, is the more durable signal: five straight quarters of roughly doubling revenue is harder to fabricate or spin than one flattering data point, and it’s the kind of growth curve that makes a September or October IPO timeline plausible rather than aspirational.

What to watch: whether either company files IPO paperwork before that window closes, and whether audited figures match what’s already been shared with investors.

More Big Tech & Industry Moves coverage, or everything tagged anthropic.

What Origin does

Origin bundles repository hosting, pull-request review and merging, and CI/deployment integrations, including Vercel, Buildkite, and Depot, into a single dashboard inside the Cursor editor, syncing with existing GitHub repositories rather than requiring a full migration off GitHub. “Origin, our code hosting platform, is now live. It’s fast, easy to use, and deeply integrated with Cursor,” the company said in its launch announcement, as reported by TechStartups.

The timing

GitHub’s outage during Origin’s launch window, which hit Actions and pull-request functionality according to that same reporting, wasn’t something Cursor could have engineered, but multiple outlets covering the launch noted the coincidence gave Origin’s GitHub-rival pitch an unusually pointed opening.

Company background

Cursor was founded in 2022 by MIT students Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger under the name Anysphere. Origin was first previewed in June 2026 at the company’s inaugural Compile conference in San Francisco, then opened to paid users in beta this week. Pricing specific to Origin hasn’t been published yet.

Why it matters

A code-hosting platform built around AI agents as first-class users, not just human developers running an AI plugin, is a different bet than GitHub Copilot’s approach of layering AI on top of an existing developer workflow. If Origin’s agent-aware review tooling proves out, it pressures GitHub to treat agent-authored code as its own category rather than folding it into standard pull request review, though it’s still early beta, limited to paid users, with no independent reliability or usage data yet beyond Cursor’s own claims.

What to watch: whether Origin exits beta with published pricing, and whether GitHub responds with its own agent-specific tooling rather than treating this as a niche competitor.

More Software News coverage, or everything tagged cursor.

The sale

The Luce “Tailor Made” is the first production chassis from Ferrari’s new electric vehicle program and, per Ferrari’s own description, the first fully electric car in the company’s history. It was displayed at the Pebble Beach Concours d’Elegance during Monterey Car Week before the sale, and RM Sotheby’s waived its buyer’s premium on the lot since all proceeds go to the Ferrari Foundation, a registered 501(c)(3) charity.

Why “polarizing”

Multiple outlets covering the auction, including CNBC, described the Luce as a polarizing design ahead of the sale. Ferrari’s shift to a fully electric model has drawn pushback from parts of its enthusiast base since the car was first previewed, the same tension that’s followed other legacy performance-car makers moving into EVs. The $40 million result suggests that reaction didn’t extend to collectors bidding on a genuine first-of-line chassis.

Why it matters

A charity auction result isn’t the same signal as retail demand; this is a single collector-grade chassis, not a market read on how the production Luce will sell once deliveries start. But it is a real data point on how Ferrari’s most serious collectors are pricing the brand’s electric transition: a record-setting bid for Chassis 0 specifically, not just any Luce, says something about how much value the market puts on being first, independent of the electric-versus-combustion debate playing out in enthusiast forums. Whether that translates into retail order volumes is the test still ahead.

What to watch: Ferrari hasn’t yet confirmed a delivery date or full pricing for the production Luce lineup beyond this one-off sale.

More Tech Automobile coverage, or everything tagged ev.