6 min read · 14 sources

Anthropic has filed its S-1 prospectus with the SEC, planning to list on the NASDAQ under the ticker ‘ANT’ as soon as the week before Thanksgiving. Bankers are targeting a public market valuation between $1.5 trillion and $2 trillion.

The 400-page filing pairs standard existential risk disclosures - ranging from bioweapons synthesis to uncontainable model autonomy - with explosive operational scaling. The financial teardown tells the real story: Anthropic generated $16.3 billion in revenue for the first half of 2026, a 14-fold surge from $1.2 billion in H1 2025. It also notched $974 million in GAAP operating income in Q2 2026 alone.

Beyond the top line, the infrastructure commitments are staggering. Anthropic is sitting on $637 billion in long-term compute obligations spread across Broadcom, SpaceXAI, Google, and AWS. The driver behind the margin expansion appears to be Jevons’ paradox in action: token cost cuts reaching 45% on Claude Fable 5.1 and Opus 5.5 catalyzed consumption, expanding its base to more than 11,900 enterprise accounts paying at least $100,000 annually.

Anthropic holds $637 billion in compute commitments across Broadcom, SpaceXAI, Google, and AWS while hitting $974 million in GAAP operating income in Q2 2026.

DoorDash moves to swallow Instacart for $17B

Source: wsj.com ↗

DoorDash is in advanced talks to acquire Instacart in a deal valuing the grocery delivery platform at roughly $17 billion. The transaction could close within days, though final terms remain fluid.

If executed, the acquisition rolls Instacart’s deep grocery catalog integrations - including enterprise relationships with Kroger, Safeway, and Aldi - into DoorDash’s dominant restaurant logistics machine. For systems operators, this is an immense data-integration problem: reconciling distinct driver dispatch graphs, varied routing constraints (hot food dispatch latency vs. multi-item picking times), and two separate retail SKU and inventory sync architectures across hundreds of chains.

CPython drops its copy-and-patch JIT

Source: lwn.net ↗

The experimental JIT compiler introduced in Python 3.13 is heading for the scrap heap. CPython maintainers have officially withdrawn standards-track PEP 836, prompting the removal of the copy-and-patch JIT architecture originally landed under the informational PEP 744.

The Python steering council had already paused active work on the main branch back in June, citing unresolved engineering friction over long-term maintenance overhead, platform compatibility, unclear success metrics, and missed milestones. Because copy-and-patch compiles C bytecode snippets into template binaries at build time and stitches them at runtime, it promised lightweight JIT benefits without a full LLVM dependency. Instead, the implementation stalled. Python developers running standard CPython runtimes will not have a production JIT compiler in the immediate future, shifting the near-term optimization burden back to tier-two Tier 2 bytecode interpretative optimizations.

Operating software factories under agent-driven commit floods

Source: claude.dev ↗

Human pull request review is collapsing under the weight of agentic workflows, forcing teams to rewrite how code enters production. At Anthropic, Claude writes approximately 90% of the code for Claude Code, which drove a radical redesign of their review pipeline. Monorepo changes spiked from 2,500 daily in May to over 30,000 in September. In response, Anthropic eliminated human line-by-line checks for agent generation, routing routine correctness and style parsing directly to evaluation models while reserving human oversight strictly for architectural boundary conditions and fast revert tooling. Merged pull requests doubled from June to September, while production revert rates dropped by over 40%.

Other engineering groups report similar architectural shifts. At Kiro, a team of just three engineers merged 1,000 pull requests in one week using an autonomous agent team. Their model reallocates 20% of core engineering hours into automated validation pipelines. Instead of reading diffs, humans audit immutable CI review logs packed with automated findings, runtime tracing, and test proofs (such as captured screenshots and automated browser interaction video) before shipping.

Steve Yegge expanded on this operational split in an essay detailing The Counting House pattern. Yegge contrasts extreme, hard-to-maintain SDLC engines like his own “Wheelhouse” - which pumps out 2,000 to 4,000 commits a week - with a more attainable enterprise blueprint. The Counting House relies on smaller, cheaper open-weight or low-tier models wrapped in rigorous, classical data science guardrails to handle enterprise workloads like support desk automation and claims triage, measuring programmatic accuracy against human baseline teams without rewriting the underlying developer stack.

Turbopuffer v3 moves to fully vectorized search

Source: turbopuffer.com ↗

Turbopuffer announced that its upcoming v3 release will feature an end-to-end vectorized query execution engine, abandoning the traditional Volcano-style row-by-row iterator pipeline common in search architectures.

Classic search engines process complex inverted-index lookups, filters, and rank scoring by passing one record at a time up an operator tree. Turbopuffer’s v3 engine borrows instead from analytical columnar engines like ClickHouse and DuckDB, executing operators over contiguous, columnar vector batches. This layout keeps CPU instruction caches saturated, maximizes SIMD register utilization, and drastically minimizes function pointer dispatch overhead. It is the first dedicated search system to make a full transition to analytical columnar query pipelining.

Infrastructure shifts: Solar power, self-hosting risks, and Instagram runtimes

Source: bbc.com ↗

Solar generation quietly surpassed nuclear power in 2025, generating 13% more total electricity than nuclear facilities across the 12 months ending in September 2026. Solar has scaled 2,700-fold since 2000, accounted for 75% of global grid demand growth in 2025, and is on pace to generate 10% of total world electricity by the end of 2026. Because levelized costs have cratered to make it the cheapest power source in history, massive data center operators are increasingly forcing utility interconnections around high-capacity solar and battery storage corridors.

The security realities of raw infrastructure maintenance are driving similar consolidations. Ben Thompson reflected on the state of the web, tracking the shift from hand-rolled PHP and MySQL setups to managed platforms like Automattic’s Pressable. The catalyst: an ubiquitous, critical PHP vulnerability permitting remote code execution on bare host machines across every version of the language. Managed platform vendors were able to hot-patch infrastructure fleet-wide prior to disclosure, a threshold that bare-metal and self-hosted environments consistently struggle to maintain as automated discovery tooling advances.

Further up the stack, Instagram is testing “Features,” a video tool supporting uploads up to 90 minutes long. Shifting away from short-form loops toward full-length programming directly challenges YouTube’s monetization grip, which has paid creators over $100 billion over four years on a 55% ad revenue share. For infrastructure teams, hosting 90-minute video files introduces dramatically heavier requirements: long-tail CDN byte-range caching, extensive HLS segmenting, multi-pass encoding profiles, and high-frequency mid-roll dynamic ad-insertion pipelines far beyond the demands of Reels.

Quick hits: OpenAI, Microsoft, and Nvidia

Source: openai.com ↗

• OpenAI announces three theoretical discoveries: OpenAI published three new scientific results across harmonic analysis, complexity theory, and quantum field theory. The lab is open-sourcing raw model outputs for the harmonic analysis result for external researchers to deconstruct, assembling a working group to analyze the computational complexity findings, and deferring to physics collaborators on releasing the quantum field theory proof materials.
• Microsoft cuts headcount via buyouts: Microsoft is rolling out voluntary buyout packages to 60,000 employees across the United States, Canada, the United Kingdom, and India. The program includes severance and career transition services; leadership expects uptake to settle below 30%. The downsizing accompanies revisions to internal review rubrics and equity compensation structures.
• Nvidia eyes agent software fees: Nvidia is floating a $2,000-per-GPU annual licensing fee for software designed to run autonomous agents across AI factories. The company already charges $4,500 per GPU annually for its standard AI Enterprise suite, leaving it currently ambiguous whether this new tier operates as an add-on or a dedicated operational runtime.