📰 Tech Trends Daily — Wednesday, October 21, 2026

Today’s Keywords: Revocation pipelines stop at the browser Data Source: HN Top 30 + Lobsters Top 25. HN is today’s snapshot (30 items, 1 hiring post excluded). Lobsters direct connection timed out (IPv6 path stalled at 7KB), falling back to Jina Reader backup link; scores, comment counts, and s/XXXXX discussion links extracted from the page’s structured output. 8 posts detailed yesterday remain on the front page and are moved to the final section with follow-ups only, without repeated commentary. Included: 41 new items + 8 continuing items = 49 items.

🔥 Today’s Focus

Today, three separate incidents across different layers delivered the same verdict: the cost of breaching is already rock bottom, while the cost of discovery and revocation remains bottlenecked at the tail end. All eight of eMudhra’s intermediate CA private keys were integers under 256—effectively unlocked. Mozilla immunized Firefox users within two hours via OneCRL, but clients with embedded trust stores like Go, Electron, and webpki-roots must await new releases before these certificates can be stripped out. Over on Tor, 333 exit nodes were held by Ebury from August through October, uncovered only when a single sysadmin scripted a pass across all ~3,000 relays and used a honeypot to watch the attackers intercept SSH via paramiko and drop malware with perl. That same week, another researcher proved that fabricating 50,000 fake statistical records was enough to get ChatGPT and over 330 domains to cite them as ground truth—even slipping into academic papers and investment bank research notes. Every link in the chain is waiting for someone else to verify first.

🔒 Security & Trust

  • eMudhra Intermediate CA Compromise — EMudhra Intermediate CA Compromise. 540 points/288 comments (HN). The private keys to eight eMudhra intermediate CA certificates were compromised. The issuing root stores encompass Mozilla, Microsoft, Apple, and Chrome, with a massive downstream of clients bundling their own trust stores. Whistleblower Wayne signed the exact same statement using all eight private keys as proof, reporting it to root programs on October 17; Firefox and Chrome blocked them the same day, and Apple followed on October 19. 💬 The most chilling detail emerged from the comments and Mozilla bug 2083449: P-384 private keys should be random integers between 1 and the group order, yet all eight were strictly under 256—every single byte except the last was all zeros. eMudhra claimed its KGC script for generating ECC issuers glitched, possibly mistakenly reusing wrap/unwrap test keys written for an HSM export workflow. Commenter galadran (Mozilla) highlighted the ecosystem gap: browsers possess online revocation pipelines like OneCRL and CRLite, whereas other clients must wait for new software releases; asveikau helpfully provided manual removal commands for FreeBSD and OpenBSD.

  • Mozilla’s Actions Following eMudhra Intermediate Certificate Compromise — Protecting Firefox users following the compromise of eMudhra intermediate certificates. 111 points/41 comments (Lobsters). The Lobsters discussion centered on engineering mechanics: one commenter explained why forks like Waterfox failed tests—versions prior to 6.7.0 froze revocation data (OneCRL, CRLite) at build time, only re-enabling network sync later; older builds will never receive new blocklists without a new release. 💬 Another comment pulled back the timeline: on October 15, an anonymous user had already commented beneath the ExplainXKCD page for the “Random Number 4” comic, yet Mozilla only received the formal report on October 17.

  • How I Got Over 300 Tor Exit Nodes Removed — How I got over 300 Tor exit nodes removed. 295 points/72 comments (HN) · 24 points/1 comment (Lobsters). On the morning of Friday, October 2, 333 exit nodes infected with Ebury 1.8.3 began coordinating an SSH man-in-the-middle attack, putting 10% of Tor’s ~3,000 exit nodes under attacker control. The author’s troubleshooting technique was crude but relentless: using exitmap to pin exit nodes, he sent SSH traffic from his own Server A to Server B, with A verifying host key fingerprints and B verifying origin IPs, scanning all 3,000 relays. Roughly 12 directly returned bogus fingerprints, while over 300 simply forwarded connections to those 12 interception servers—forwarding alone was proof enough of guilt. 💬 The post-mortem’s final section is the most valuable: he deployed sshesame as a honeypot to lure attackers into logging in, discovering the interception backend was built with paramiko; the dropper script probed id -u and sudo privileges before base64-decoding a payload handed over to perl for execution. As for the initial vector, only rumors of “compromised hosting providers” circulate, with none yet confirmed.

  • Do Minor curl Vulnerabilities in the Open — Do minor curl vulnerabilities in the open. 11 points/0 comments (Lobsters). Daniel Stenberg explores disclosure boundaries for low-severity vulnerabilities: issuing advisories for every trivial issue gets dismissed as noise, but withholding them amounts to concealing risks from downstream users. Fresh on the front page, no discussion yet.

  • Upcoming Critical PHP Security Release — Upcoming Critical PHP Security Release. 8 points/1 comment (Lobsters). The PHP team pre-announced a critical security release without disclosing details—traditionally indicating a remotely exploitable memory corruption bug.

  • Setting Up SSH Keys on macOS with Touch ID — Setting up SSH keys on macOS with Touch ID. 48 points/21 comments (Lobsters). Stashing private keys inside the Secure Enclave and requiring fingerprint verification for signing operations. Paired with the Tor retrospective above, the lesson is clear: attackers deliberately target hosts still relying on passwords and plaintext private keys.

  • Feds Suspect Ransom Negotiator Ran ShinyHunters Extortion Site — Feds Suspect Ransom Negotiator Ran ShinyHunters Extortion Site. Fresh on the front page, no discussion yet (HN). Krebs’s lead points to the treacherous scenario where “the negotiator hired to help you is running the extortion ring”—a development worth monitoring closely.

🤖 AI & Models

  • DeepInfer: DeepSeek’s Production Inference Framework Open-Sourced — DeepInfer: High-performance inference framework for DeepSeek models. 2074 points/811 comments (HN). The repository description confirms this is the exact engine running DeepSeek’s massive online production services, with its offline inference pipeline now open-sourced. It currently targets DeepSeek-V4.1-Flash and packs in proprietary kernels like FlashMLA, DeepGEMM, DeepEP, TileKernels, and DeepSelect. It leverages EP + EPLB for MoE load balancing, DSpark for draft token speculative decoding with confidence-based scheduling, SWA bounded replay, batch-invariant inference, and deterministic sampling. 💬 Two details stood out: first, its support for Huawei Ascend 950 NPUs, which yaoke259 argued shifts the hardware landscape more than the model itself—allowing smaller labs to build hardware capacity around this stack; second, mmaunder highlighted the timeline: US DHS Exclusion and Removal Order #2026-01 was dated July 24 and only unsealed yesterday, suggesting DeepSeek anticipated the move well in advance and timed the open-source release accordingly.

  • Voxtral 3: Voice Agent SDK and Speech Model Supporting 50+ Languages — Voxtral 3. 471 points/130 comments (HN). Mistral packages transcription, emotion-adaptive speech synthesis, and a real-time voice agent SDK into a single bundle, offered via EU-hosted APIs or on-prem deployment. The pitch centers on data sovereignty and tool calling: in the demo, a French patient calls in, and the agent parses a referral letter, verifies insurance numbers, checks schedule openings, books the appointment, and confirms details—all within a single continuous phone call. 💬 Two points of skepticism: ComputerGuru compared audio samples on the site, noting the new generation sounds noticeably more synthetic than Voxtral 1, alongside complaints about Whisper (noting it is a transcription model, not a dictation model); ilaksh pressed on how Voxtral Large’s “human handoff” is triggered—whether via special tokens or command execution—which remains unspecified in official documentation.

  • SlopCodeBench October Update: High-Reasoning Tiers Disproven — SlopCodeBench 10/2026 – Opus 5.5, Astra wins overall, xhigh considered harmful. 124 points/50 comments (HN). Across 36 coding problems and 196 checkpoints per question over ten full campaigns, strict pass rates landed between 12.2%–16.3%, with core pass rates between 65.8%–74.0%. The real revelation is that ultra-high reasoning tiers offer terrible economics: XHigh costs 2.11x–4.22x more than Medium, yet Opus 5.5 passed 28 on Medium and only 25 on XHigh; Astra XHigh solved 32, leading SOL 6.1 Medium by just two checkpoints. 💬 Comments noted that Astra writes prose well but habitually drops newlines in code (as previously documented by lucumr), while others warned that benchmaxing turns models into rote test-takers: the more granular the metric, the more the model sacrifices broad engineering judgment just to pump benchmark scores.

  • Claude Designed Protein Binders Against Scorpion Neurotoxins — Claude designed protein binders against scorpion neurotoxins. 80 points/42 comments (HN). The challenge was posed by Citrini Research: could motivated non-specialists tackle a real-world public health crisis? Claude ran for 64 hours across one prompt and four follow-ups; Adaptyv Bio and Twist Bioscience conducted wet-lab validation on 211 designs, yielding 17 hits, with all 211 designs and datasets released publicly. Context: scorpion stings affect 1.2 million people annually, killing over 3,250—mostly children—yet poor commercial incentives have long relegated it to neglected tropical disease status. 💬 Commenters offered mixed reactions: one noted that the team using RFdiffusion in 2025 for snake antivenom already achieved in vivo mouse survival validation, representing a more advanced milestone; another claimed to have run the same workflow at home on an RTX 5090 against other targets, arguing BindCraft2 is a better starting point—calling it “not exceptional, just tool orchestration,” though acknowledging the scale of screening 318,000 designs is impressive.

  • Google Open-Sources TorchTPU: Native PyTorch Back End for TPUs — Google open-sources TorchTPU, a native PyTorch back end for TPUs. 223 points/53 comments (HN). Integrating TPUs directly into PyTorch’s official backend abstraction effectively opens up the XLA compilation stack straight from the PyTorch side. 💬 An intriguing crossover came from pcmoritz, who is writing an open-source libtpu implementation (libtt) for Tenstorrent accelerator cards, noting most TorchTPU use cases already run; another reminded that AWS has taken this approach with Trainium for years (torch-neuronx). Others saw this as the final nail in TensorFlow’s coffin.

  • Publish Fake Stats, Get Cited in LLMs — Publish fake stats, get cited in LLMs. 307 points/176 comments (HN). An author seeded a website with 50,000 completely fabricated statistical data points: Google delisted it, but ChatGPT did not. Over seven months, it drove over 10,000 AI conversations and was cited by more than 330 domains, including Shopify, Bluehost, and DHL, with fictitious numbers leaking into academic literature and bank investment reports. 💬 Commenters tied this to established dynamics: citogenesis (circular reporting on Wikipedia) has existed for decades; models simply hyper-accelerate the cycle. Others noted an even lazier exploit: writing fake reviews for your own product, which models ingest and regurgitate as authoritative evidence of customer satisfaction.

  • Launch HN: Respan (YC W24) — Open-Source Rust AI Gateway Built Around Plugins — Launch HN: Respan (YC W24) – Open-source Rust AI gateway built around plugins. 36 points/16 comments (HN). The core differentiator lies in extension points: request rewriting, routing, billing, and logging are all engineered as plugins rather than config directives. Competition at the gateway tier has shifted toward plugin ecosystems and debuggability.

🛠️ Tools & Infrastructure

  • Inside Prometheus: Meta’s Gigawatt-Scale Supercluster — Meta’s GB{2,3}00 Cluster. 44 points/20 comments (HN). Meta’s New Albany, Ohio campus—serving standard web traffic just two years ago—has been transformed into a frontier training ground. Phase one is operational, training the Muse family, with an ultimate target of ~1 million GPUs utilizing closed-loop liquid cooling. The Ariel variant of GB200 NVL72 pairs every Blackwell GPU with one Grace CPU and two ConnectX-7 400G NICs, achieving a 1:1 CPU-to-GPU ratio and 800G I/O per GPU. 💬 The comment section was far from celebratory: one user noted Meta’s crawler accounted for 90% of requests to their server over a week, completely ignoring robots.txt and persisting despite IP bans; another pointed out the post was published two days before earnings, suspecting bad news on capital expenditure.

  • Google Cloud Outage Affects Almost All Banking and Other Operations in Europe — Google Cloud outage affects almost all banking and other operations in Europe. 302 points/156 comments (HN). Knocking out operations across HSBC, Revolut, Bank of Scotland, Lloyds, and Equifax, the incident stemmed from a power event in the europe-west2-b zone according to Google’s status dashboard. 💬 The most visceral takeaway came from a Revolut customer: primary accounts unreachable, no cash withdrawals, cards declined, no phone lines, and no staff to talk to—the dark underbelly of pure neobanks. Another noted that Lloyds’s login issues would linger long after the underlying cloud recovered.

  • Tsrs: A Port of TypeScript to Rust — Tsrs: A Port of TypeScript to Rust. 322 points/146 comments (HN). Built jointly by Cloudflare, VoidZero, and Max Schwenk of Owner.com within the Oxc toolchain ecosystem, early benchmarks show up to 8x faster performance and ~70% lower memory usage under identical type-check worker configurations. The team maintains active collaboration with the official TypeScript team, explicitly choosing not to fork. 💬 Two explanations added crucial context: one pointed out their bespoke memory management model (detailed in their PORTING doc), using “never free” arena-style allocation to sidestep the structural hurdles of porting GC languages to Rust; another revealed the underlying motive—Cloudflare needs tsc running inside pure WASM sandboxes, whereas a Go port would drag in the entire Go runtime; raw speed was merely a welcome side effect.

  • Jepsen: Orkes Conductor 5.3.2 — Jepsen: Orkes Conductor 5.3.2. 48 points/1 comment (HN) · 26 points/1 comment (Lobsters). Kyle Kingsbury evaluated versions 5.3.2 through 5.5.0-rc2: tasks experienced concurrent duplicate execution more frequently than expected, alongside three critical flaws causing workflows to be permanently abandoned or forgotten, or tasks marked complete without ever running. Version 5.5.6 has patched all critical security issues and mitigated concurrent duplicate runs. The analysis was funded by Orkes.

  • MicroG: Free-Libre Re-Implementation of Google’s Proprietary Android User Space — MicroG: Free-libre re-implementation of Google’s proprietary Android user space. 369 points/85 comments (HN). A veteran project resurfacing on the charts, allowing GMS-free devices to run apps reliant on Google Play Services. 💬 The discussion centered on its boundaries: one commenter cited France’s SNCF Connect app, puzzled why MicroG users could process payments while their own setup failed; others noted it fundamentally relies on signature spoofing—while scoped narrowly on LineageOS, it still presents an expanded attack surface.

  • IKEA – 3D Guide — IKEA – 3D Guide. 122 points/38 comments (HN). Swapping flat paper exploded diagrams for interactive, rotatable 3D assembly steps—neatly solving the eternal paper manual frustration of determining which side of a board faces which way.

💻 Languages, Compilers & Systems

  • Why Do We Need ud2 When int 6 Is Right There? — Why do we need ud2 when int 6 is right there?. 10 points/0 comments (Lobsters). Raymond Chen explains why x86 introduced dedicated ud2 instructions for “code that should never execute” instead of reusing the generic invalid opcode exception (int 6). Such architectural trivia illuminates why modern toolchains can reliably depend on an unambiguously defined fallback instruction.

  • A Very Fast IPv4 Parser, in Swift — A Very Fast IPv4 Parser, in Swift. 17 points/1 comment (Lobsters). Parsing a.b.c.d sits on an exhaustively hot path; the author re-implemented it from scratch in Swift alongside thorough benchmarks. Discussions around Swift’s systems performance outside Apple ecosystems remain lively.

  • leanin: A Tokio-Inspired Concurrency Library for Lean 4, with Machine-Checked Proofs — leanin: A Tokio-inspired concurrency library for Lean 4, with machine-checked proofs. 1 point/0 comments (Lobsters). Formalizing runtime concurrency primitives alongside their verified invariants, paving the way toward turning “this async code deadlocks” into a provably falsifiable theorem. Newly published, discussions haven’t kicked off yet.

  • Declarative WASM GC Types — Declarative WASM GC Types. 2 points/0 comments (Lobsters). Analyzing the architectural gains of transitioning the WebAssembly GC type system from structural recursive definitions to declarative specifications—a frontier topic in type theory circles.

  • Making a VGA Driver in SystemVerilog — Making a VGA Driver in SystemVerilog. 16 points/9 comments (Lobsters). Writing pixel clocks, horizontal and vertical sync pulses from scratch up to active display rendering without relying on off-the-shelf IP. The cleanest bottom-up path to demystifying display timing requirements.

  • EDG Is Open Source — EDG is open source. 3 points/4 comments (Lobsters). Edison Design Group’s C++ frontend is legendary for standards compliance and diagnostic quality, having operated under proprietary licensing while serving as a reference benchmark for compiler vendors. The broader toolchain impact hinges on license specifics.

  • OpenBSD 8.0 — OpenBSD 8.0. 94 points/10 comments (Lobsters). Continuing from yesterday’s Wine port discussions, this release notably brings native response header modification support to httpd—a feature previously requiring a fronting relayd proxy. 💬 Several users had already been tracking snapshots specifically for this capability and can now run it on stable.

  • NixOS Is Making Me Less Afraid — NixOS is making me less afraid. 35 points/22 comments (Lobsters). The author reflects on how fear of system tinkering stems from irreversibility; declarative configurations paired with atomic rollbacks turn “let me try this” into a safe, reversible action. The resonance is emotional rather than purely technical.

🌐 Platforms, Policy & Ecosystem

  • Facebook Pages Require Paid Subscription to Post More Than Two Links per Month — Facebook Pages require paid subscription to post more than two links per month. 397 points/357 comments (HN). Developer reports reveal Meta is capping outbound links for Facebook Pages to just two per month, forcing paid subscriptions for anything more. 💬 Commenters provided sharp historical context: in the 2010s, businesses were urged to “replace websites with Facebook Pages,” only for Meta to downrank Page posts in newsfeeds; this latest move escalates that trajectory. One comment captured it succinctly: the 2020s are a war on the hyperlink, cloaked under the guise of “external link safety.”

  • Ofcom Concedes Defeat in 4chan Case — Ofcom concedes defeat in 4chan case. 293 points/202 comments (HN). Ofcom CEO Dame Melanie Dawes admitted before the House of Commons DCMS Committee that fines cannot be collected when foreign courts stand in the way. Author Preston Byrne, an attorney representing 4chan in the case, cited earlier rulings from US courts in 4chan v. Ofcom: because the entity maintained no physical presence, infrastructure, or operations in the UK, enforcement fell to US courts, which routinely refuse to enforce foreign judgments violating First Amendment rights. 💬 Comments split in two: one camp highlighted that cross-border enforcement toolkits are inherently constrained—threats of border arrest and payment processor blacklisting are failing; the other feared the UK will pivot to DNS/network-level blocking, replacing “regulating abroad” with “severing at home.”

  • Apple Broke an Essential Presentation Feature of iPads — Apple broke an essential presentation feature of iPads. 457 points/243 comments (HN). Written by a university lecturer whose workflow relied on handwriting notes on an iPad, projecting them to students via AirPlay, while using the same device to drive Mac slide decks, check rosters, and review poll results. Starting with iPadOS 27, tapping any non-presentation app instantly switches the external display to full-screen mirroring—exposing rosters and quiz keys to students. 💬 Fellow educator pflats conceded the legacy interaction was obtuse (requiring swiping to the home screen and selecting apps from the dock to take over the external screen), but breaking it is not fixing it; prompting a modal choice between “Mirror Entire Screen” or “Keep Current Window” would be far saner. Others lamented the decline of iPad window management since iOS 26.

  • Mastodon 5.0: A Closer Look — Mastodon 5.0: A closer look. 14 points/3 comments (Lobsters). The release notes focus primarily on UI revamps and quote-post mechanics. Recent Fediverse attention has largely drifted toward client ecosystems rather than core protocol shifts.

  • Academics Have a Duty to Defend Open Knowledge Systems Like Wikipedia — Academics have a duty to defend open knowledge systems like Wikipedia. 48 points/20 comments (HN). An LSE commentary arguing scholars must actively edit entries, fact-check claims, and push back against automated model-generated sludge. Directly mirrors the “publishing fake stats” report: open knowledge repositories rely on humans, and humans are being crowded out by low-friction automation.

🔬 Science & Other

  • Sixteen Years After Deepwater Horizon, Deep Sea Life Shows Few Signs of Recovery — Sixteen Years After Deepwater Horizon, Deep Sea Life Shows Few Signs of Recovery. 141 points/95 comments (HN). Researchers cautiously stopped short of directly attributing population collapses solely to the spill, sparking debate in the comments. 💬 One commenter offered scale: the 1979 Ixtoc I disaster spilled 3.3 million barrels (compared to Deepwater Horizon’s 4.9 million), and historical findings already established that exposed, high-energy shallow waters recover, while sheltered deep-sea environments do not. Another brought up the nearly 2 million gallons of chemical dispersants used, which made oil denser and forced it to the seafloor—hiding the catastrophe from public view.

  • JAXA Is About to Perform a Phobos Sample Return Mission — JAXA is about to perform a Phobos sample return mission. 80 points/14 comments (HN). Scooping regolith from Phobos and returning it to Earth represents a far more aggressive posture than the Hayabusa missions: the moon’s minuscule gravity requires redesigning both sampling mechanisms and ascent trajectories from the ground up.

  • Making Solar Panels with a Microwave — Making Solar Panels with a Microwave [video]. 18 points/2 comments (HN). A hands-on home lab video hacking a consumer microwave into the silicon wafer treatment process. Commercial efficiency is out of reach, but the real draw is seeing how drastically the fabrication pipeline can be compressed.

  • Silicon-Photonics-Enabled Chip-Based 3D Printer (2024) — Silicon-photonics-enabled chip-based 3D printer (2024). 4 points/0 comments (HN). A resurfaced 2024 paper integrating stereolithography optical pathways directly into silicon chips, shrinking the curing apparatus to fingertip scale. A quiet comment section suggests it is still searching for its production audience.

🎮 Light & Fun

  • SimCity, but with Permits — SimCity, but with Permits. 419 points/226 comments (HN). In Permit Pending, you are not an omnipotent deity; you are an elected mayor. Every apartment complex, bus lane, and factory must crawl through hearings, reconsiderations, lawsuits, and appeals; developers and NIMBYs alike are completely unreasonable, and the underlying macroeconomic simulator is unsparing. The homepage sets expectations upfront: “Housing prices climb another month. Youth blame you for lack of supply; homeowners sue you over reckless approvals; the election is in 48 months.” 💬 Half the comments filed bug reports (WebGL texture allocation overflows turning screens black, and new games crashing with PP-SIM-F99D), while the other half dissected why permitting structures inevitably stifle good projects. bruce511 gave the sharpest breakdown: the rules are derived from trying to please everyone at once, making noise ordinances, parking minimums, and low-density mandates mathematically inevitable.

  • What Shall I Teach My Son? — What shall I teach my son?. 193 points/267 comments (HN). Prompted by the realization that “programming as a craft is depreciating,” the author ponders what enduring skills to pass on. 💬 The top comment pushed back against framing parenting as a technical question: what parents pass on is rarely hard technical skills, but curiosity, agency, and the hours spent reading together, hiking, and exploring grocery markets. Another corrected the essay’s tropes—Bill Gates excelled academically at Harvard and published co-authored papers; using him as rhetorical evidence that “dropping out leads to success” is fundamentally flawed.

  • ‘Der Viener Schlinger’: The Meat Cannon That Launched 20k Hot Dogs — ‘Der Viener Schlinger’: The meat cannon that launched 20k hot dogs. 84 points/49 comments (HN). Local investigative journalism tracking grassroots engineering history: a homemade pneumatic compressed-air cannon that grew into a regional sports fixture.

  • My Friend Bob Cringely — My Friend Bob Cringely. 150 points/39 comments (HN). A tech journalist’s personal reminiscence exploring how columnists built trust in that era, and how it eventually unraveled. 💬 Discussion leaned nostalgic, with many remembering reading his columns before knowing the author, inevitably leading to questions about whether the personal column format has any viable future left.

📌 Yesterday’s Trends, Follow-ups Only

📝 Summary

Today’s focus rests on the final links of the trust chain: while browser vendors have industrialized revocation pipelines, non-browser clients remain stranded on manual release schedules, and detection still leans heavily on individual dedication—333 infected Tor exit nodes were not flagged by monitoring systems, but uncovered by one engineer scanning every relay across the entire network. Three essential reads: the eMudhra incident page (why eight private keys were effectively zero, and which clients remain vulnerable), the Tor exit node triage post-mortem (how crude brute-force scans audited ~3,000 relays and what honeypots revealed about the dropper scripts), and the DeepInfer repository (DeepSeek open-sourcing its production inference engine alongside Huawei Ascend NPU support, timed immediately following the public unsealing of a DHS removal order). The cross-cutting signal is that retreat and substitution are unfolding simultaneously: the Netherlands is building its sovereign Linux distribution, Claude-designed scorpion venom binders cleared wet-lab validation with 211 published designs, Cloudflare ported tsc to Rust to slash memory by ~70%, and Meta is scaling gigawatt clusters toward a million GPUs—critical dependencies are being pulled back into internal control across the board, each for distinct reasons.