Daily2026-08-22
TechEconomic

Nvidia writes a $6 billion license check to Poolside, hires 109 engineers, invests another $1 billion at a $12 billion mark

The deal Nvidia announced Friday buys a non-exclusive license to Poolside's Model Factory training stack and pulls the 109 engineers behind the Laguna family of open-weight coding models onto Nvidia's payroll, while leaving Poolside standing as an independent company with a fresh $1 billion of Nvidia equity capital and three co-founders in place. The template mirrors the Groq licensing structure from earlier this year and is Nvidia's fourth in twelve months, ratifying licensing plus talent transfer as the deal shape frontier-model shops now take instead of an acquihire.

PYMNTS · Newcomer · Bloomberg · The Information · The Decoder · 2026-08-21 · Source ↗

Nvidia agreed Friday to pay Poolside about $6 billion for a non-exclusive license to the software Poolside uses to build its Laguna family of open-weight coding models, and made a separate $1 billion equity investment at a $12 billion pre-money valuation. The company will also extend job offers to 109 Poolside engineers involved in Laguna development. The three Poolside co-founders stay in place and the company continues to operate independently. Newcomer's Eric Newcomer surfaced the letter to investors describing the terms, and PYMNTS and Bloomberg confirmed the structure on Friday.

"Poolside AI, the artificial intelligence model-building startup, has struck a non-exclusive licensing deal with Nvidia for $6 billion, plus a $1 billion investment in Poolside at a $12 billion pre-money valuation, according to a letter to investors obtained by Newcomer." — Eric Newcomer, Newcomer

What Nvidia actually buys

The license covers Poolside's Model Factory, the internal training and RL stack that produces the Laguna open-weight coding models. Model Factory is the piece of software that turns a pile of GPUs and coding-benchmark data into a working code-completion model, and Poolside built the stack to run at frontier scale on Nvidia hardware. The 109 engineering offers cover the researchers and infrastructure people who ran that stack in production. Nvidia has been telegraphing for a year that it wants to become a first-class open-weight model shop of its own, and Laguna is the missing piece: a competitive coding model whose weights Nvidia can publish and whose training loop it now owns.

The deal is not an acquihire

Poolside continues to exist as an independent company. The $1 billion Nvidia equity check funds the residual business at a $12 billion pre-money valuation, giving Poolside dry powder to keep pursuing product lines outside coding and to recruit a fresh engineering roster. The three co-founders remain. Multiple outlets emphasized this point Friday because it distinguishes the Poolside arrangement from a Meta-style acquihire, where the target shell is wound down and the technology is absorbed. Poolside stated in the investor letter that it lacked sufficient Nvidia hardware access to continue competing head-on in open-source model development independently, which is the specific constraint the license-plus-investment structure solves.

"The reported $6 billion fee is expected to be distributed to Poolside's investors by the end of 2027." — PYMNTS, summarizing the payout schedule

Fourth time Nvidia has run this play

The Poolside terms almost exactly match the deal Nvidia struck with Groq earlier this year: a roughly $20 billion non-exclusive license for Groq's LPU architecture combined with an equity check that left Groq operating on its own, plus talent transfers into what Nvidia branded as the "Nvidia Groq 3" LPU line. Two other Nvidia deals in the past twelve months have taken the same shape at smaller scale. Antitrust regulators have been paying attention. The Justice Department's ongoing look at Nvidia's dominance in AI training silicon has become sharper with each of these structures, since a license-plus-talent play sidesteps the merger-review threshold that an outright acquisition would trigger while producing effectively the same market outcome.

Poolside's runway and what happens to CoreWeave

Poolside spent early 2026 building a $10 billion sales-and-lease-back platform with CoreWeave for compute capacity, and the Nvidia license does not by itself unwind that arrangement. What the deal does do is give Nvidia line-of-sight into Poolside's remaining coding work while transferring the core Laguna engineering back to Nvidia's own R&D roster. The Model Factory license is non-exclusive, which means Poolside is free to keep using its own stack for other products. In practice, Poolside is now more customer than competitor to Nvidia's open-weight coding strategy, and the equity investment aligns its incentives with Nvidia's for the next several years.

What the market read into it

Nvidia stock closed the week down 5 percent, per Stocktwits' end-of-week roundup, on concerns that the Poolside price plus a rumored $7 billion combined value on prior similar deals is compressing Nvidia's model-line ROIC. The counter-argument, which Nvidia bulls made throughout Friday's trading, is that Nvidia is buying a stable of open-weight models it can bundle with silicon at negligible marginal cost, and that the total license cost is a rounding error against the company's now-$95 billion quarterly revenue trajectory heading into the Q2 FY27 print due Tuesday, August 26.

Big Picture

Friday's tape lined up three separate signals that the AI capital cycle is entering a new phase. Nvidia paid Poolside $6 billion for a Model Factory license and hired 109 of its engineers, the fourth licensing-plus-talent deal Nvidia has closed in twelve months and the strongest read yet on how the biggest chipmaker plans to grow into an open-weight model shop without triggering a merger review. Bloomberg confirmed that Anthropic expects its IPO to match or top SpaceX's $75 billion record, with the S-1 filing targeted for as soon as the end of August, Citigroup added to the underwriter roster on Friday, and the deal now anchored on run-rate revenue that crossed $65 billion at the end of July. Jensen Huang met Rebellions co-founder Sunghyun Park at Nvidia's Santa Clara headquarters this week to open early talks on a partnership, investment, or acquisition of the Bundang-based NPU designer valued around $2.3 billion, extending the same licensing-heavy playbook into inference silicon.

Product news skewed toward AI reaching into surfaces users already have open. OpenAI shipped Codex Remote to general availability across every ChatGPT plan, added ChatGPT Voice inside the Codex desktop client, and let Site owners on Plus and Pro change a Site's hosted URL without redeploying, while xAI took Grok Bot out of beta and made the AI teammate available on SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and Cursor Teams. Google DeepMind's Gemma family crossed one billion cumulative downloads with more than 100,000 community variants published, and Apple Music sent partners a Made With AI labeling framework that will roll out to listeners later this year, requiring distributors to attach an AI Transparency Tag whenever generative AI is used to create a material portion of a track. Twin1 launched from stealth with $20 million to build AI digital twins of individual knowledge workers, and Astromech, the Colossal spinout run by Ben Lamm and George Church, hit a $3.8 billion valuation on a fresh $20 million round for a biological forecasting model.

Policy and safety kept up the pace. Nevada Transportation Authority commissioners unanimously approved commercial robotaxi permits for Tesla, Waymo, and Uber covering up to 8,000 vehicles across Clark County over the next twelve months, the single largest robotaxi authorization on the books. Judge Vince Chhabria tossed the seven economic-espionage counts against former Google engineer Linwei Ding while leaving the seven trade-secret theft convictions intact ahead of a September 1 sentencing. Super Micro's board investigation cleared current senior management on the $2.5 billion Nvidia-chip diversion scheme but fired several employees. Brazil unveiled a $444 million AI investment plan bringing US and Chinese partners to the same table, and CBRE's annual survey put New York's 394,300 tech jobs ahead of San Francisco's 375,730 for the first time in the report's history, driven by AI hiring on the finance side of the economy. Hugging Face's Papers of the Day for August 21 clustered around embodied and generative work, led by Ant Research's 4DAnyone monocular-to-4D reconstruction paper at 162 upvotes.

More today

17 stories · 2026-08-21
Economic

Anthropic expects to match or top SpaceX's record $75 billion IPO

Bloomberg · Katherine Doherty · 2026-08-21 · Source ↗

Anthropic is running valuation scenarios ahead of a potential S-1 filing "as soon as the end of this month," per Bloomberg reporters citing people familiar with the deal. Citigroup joined the underwriter roster on Friday, adding to Morgan Stanley, Goldman Sachs and JPMorgan already in the mandate, with pricing anchored on run-rate revenue that crossed $65 billion by the end of July.

Anthropic expects its planned IPO to match or exceed the size of SpaceX's $75 billion record listing, Bloomberg reported Friday. The Claude developer is stress-testing valuation ranges ahead of a public S-1 filing that could land as soon as the end of August. Citigroup is being added to the bank syndicate, joining Morgan Stanley, Goldman Sachs and JPMorgan, all named on the mandate earlier this month. The push is being driven by what people familiar with the discussions described as overwhelming demand from institutional investors chasing AI exposure that has been effectively locked up in private hands since Anthropic's May Series I.

"Anthropic Expects to Match or Top SpaceX's Record IPO Size." — Bloomberg headline, August 21, 2026

The revenue base under the deal

Anthropic's annualized revenue reached about $65 billion by end of July, per the Bloomberg-reviewed materials, up sharply from the $11.6 billion Q2 print Fortune surfaced last week. That growth rate is the specific piece of the story that makes a $75-billion-plus IPO valuation defensible to public-market allocators who have watched Anthropic's revenue ramp compress against OpenAI's own $40 billion enterprise annualized rate. The July run-rate implies gross annualized revenue in the third quarter closer to $80-90 billion if the current pace holds.

Timing and syndicate

The end-of-August S-1 target is aggressive. Getting the working group aligned on valuation, drafting the risk factors, and completing the SEC comment cycle inside two weeks is possible but tight; the deal could slip into September without changing the overall arc. Citi's addition to Morgan Stanley, Goldman and JPMorgan gives Anthropic a fourth relationship bank on the printer, unusual for a first-time issuer, and one investor Bloomberg spoke to interpreted the added slot as Anthropic distributing the fee pool to keep everyone motivated on distribution. Two other banks Bloomberg did not name are said to be in discussions for co-manager roles.

"Wall Street jostles for roles." — Bloomberg's read on the underwriter competition

Class structure

Founders Dario and Daniela Amodei are preparing a supervoting class ahead of the listing, per The Information's reporting from Tuesday. The dual-class structure would concentrate governance rights with the founders through a decade-long sunset, mirroring what Meta and Alphabet did at their own IPOs. Anthropic's Series I in May added $30 billion at a $900 billion post-money on the primary and gave the founders room to run a public deal without diluting operational control.

TechEconomic

Nvidia opens early talks with Korean inference chip startup Rebellions

Bloomberg · Sohee Kim, Ian King · The Next Web · Seeking Alpha · 2026-08-21 · Source ↗

Jensen Huang met Rebellions co-founder and CEO Sunghyun Park at Nvidia's Santa Clara campus this week to discuss a partnership, investment, or outright acquisition of the Bundang-based NPU designer, per Bloomberg. Rebellions raised more than $400 million in March at pre-IPO terms, sits at a $2.3 billion valuation, and has taken $850 million total from SK Hynix, Samsung Ventures, Arm, and the Korean government.

Nvidia is in early talks with Rebellions Inc., the South Korean AI chip startup, on a potential collaboration that could range from a strategic partnership to a full acquisition, Bloomberg reported Friday. Jensen Huang met Rebellions co-founder and CEO Sunghyun Park at Nvidia's Santa Clara headquarters this week to open the conversation. No price, structure, or stake size has been reported. The talks are described as preliminary and may not lead to a final deal.

What Rebellions makes

Rebellions, based in Bundang south of Seoul, designs neural processing units optimized for AI inference workloads in data centers. The company's Atom and Atom Max NPUs went into mass production in 2023 and are deployed by customers in Japan, Saudi Arabia and the United States. The design philosophy targets the inference tier of the AI workload, which is less computationally intense than training but represents the higher-volume run-rate demand that a production AI service actually generates. Rebellions merged with Sapeon Korea in 2024 and last year partnered with Marvell on sovereign AI hardware for markets that require domestic silicon supply.

Cap table and valuation

Rebellions has raised approximately $850 million cumulatively from SK Hynix, Samsung Ventures and Arm, alongside direct backing from the South Korean government. A round in March raised more than $400 million on pre-IPO terms and brought the valuation to about $2.3 billion. SK Hynix's investor position is the specific piece that will complicate any Nvidia deal, because SK Hynix is Nvidia's largest HBM supplier, and any transaction that consolidates NPU design and HBM supply under a single strategic umbrella will draw regulatory attention from both the US Department of Justice and Korea's Fair Trade Commission.

"Any transaction would navigate complex regulatory waters, as Nvidia's towering market share draws close attention from antitrust regulators like the US Department of Justice, and South Korea regards advanced semiconductors as critical national assets." — The Next Web, on the regulatory posture

Fourth Nvidia deal of this shape in a year

The Rebellions conversation is the fourth licensing-heavy or absorb-the-talent play Nvidia has opened in the past twelve months. The Poolside license announced the same day is the most recent, and the Groq LPU deal earlier in 2026 supplied the structural template: pay a large non-exclusive license fee, hire the core engineering team, take an equity position, leave the target operating independently. If the Rebellions conversation converges on that shape, it gives Nvidia a Korean inference-silicon design house without triggering a merger review in either jurisdiction. The trade-off is that Korea has been unusually protective of strategic semiconductor assets under the current administration and would likely require domestic-supply guarantees on any consummated deal.

Tech

Google's Gemma open-weight family crosses 1 billion downloads and 100,000 community variants

Unite.AI · Google DeepMind · BigGo · 2026-08-21 · Source ↗

The Gemma open-weights family reached one billion cumulative downloads in the two years since first release, with more than 100,000 community variants published by outside developers. Clement Farabet and Olivier Lacombe of Google DeepMind called out live deployments at NASA, satellite startup Satlyt, orbital-compute company Starcloud, and the Yale/Google C2S-Scale team that identified a novel cancer therapy pathway subsequently confirmed in living cells.

Google DeepMind's Gemma family of open-weight models passed one billion cumulative downloads, the company disclosed on Friday. The count covers all Gemma variants Google has released since February 2024, including the original Gemma text models, Gemma 2, PaliGemma vision-language, CodeGemma for coding, DataGemma for structured data, ShieldGemma safety classifiers, TxGemma for therapeutics research, and the Gemma 3 family that closed 2025. Outside developers have published more than 100,000 distinct Gemma variants, including fine-tunes and derivatives adapted to specific languages, tasks, and hardware targets.

"Gemma at one billion downloads is a milestone for open, accessible AI." — Clement Farabet, VP, Google DeepMind

Where Gemma actually runs

Google spent the announcement pointing at concrete deployments rather than benchmark rankings. NASA, satellite startup Satlyt, and orbital-compute company Starcloud are running Gemma models on-satellite for image analysis, deciding what is worth sending down over scarce downlink bandwidth and routing communications between satellites. A Yale-Google collaboration on C2S-Scale, an AI model built on Gemma, discovered a novel cancer therapy pathway that was subsequently verified in living cells. Both examples are Google's argument for why the open-weights strategy matters: workloads that can't reach a cloud API in real time or that require training on proprietary data can still take advantage of a frontier-competitive model when the weights are open.

Where Gemma sits against the competition

Gemma remains second-tier by raw benchmark performance against DeepSeek, Qwen, and GLM, but the download milestone is the deployment-scale answer to that. Meta's Llama 4 family and DeepSeek V4-Pro sit ahead on capability leaderboards; Alibaba's Qwen family is ahead on cumulative downloads at more than 3 billion per the Hugging Face state-of-open-models report last weekend. Gemma's differentiator has been vertical variants (medical, code, safety, therapeutics) shipped as first-class members of the family, and Google is using the download crossing to defend the strategy against outside pressure to consolidate on fewer, larger models.

Stock reaction

Alphabet closed lower Friday despite the milestone, roughly 1.8 percent, in a broader semiconductor and mega-cap tech pullback tied to the Broadcom-Marvell repositioning earlier in the week. The Gemma news landed as an infographic post on the Google DeepMind blog Thursday afternoon and had circulated through the developer press by Friday morning.

TechApplication

xAI takes Grok Bot to general availability across SuperGrok Plus, Heavy, and Cursor

Digital Applied · xAI · 2026-08-21 · Source ↗

Grok Bot moved out of beta Friday after ten days of testing, becoming available to SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and Cursor Teams subscribers. Each bot runs its own always-on cloud computer, handles work across email, calendars, docs, and code, and can be invoked in Slack, X threads, and Cursor sessions using @-mention conventions.

xAI made Grok Bot generally available on Friday, ten days after the August 11 beta launch. The product is an always-on AI teammate that runs on its own cloud VM, holds persistent state across conversations, and can be summoned into Slack channels, X direct messages, Cursor sessions, and email threads through @-mention triggers. The GA release opens Grok Bot to SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and Cursor Teams subscribers. Free-tier Grok users do not get bot access at GA; xAI has said it will evaluate broader rollout after Q3.

What Grok Bot actually does

Each bot is provisioned as a lightweight VM with Grok 4.6 as its default model, an isolated file system, a persistent memory store, and connectors to a user-authorized set of external tools. The tool set at GA includes Gmail, Google Calendar, Google Drive, Slack, X, GitHub, Cursor's Origin git platform, and a general web-browsing surface backed by xAI's Colossus data. Users can hand a bot a long-running objective (draft a specification, monitor a set of PRs, track a topic across news sources) and check back later; the bot serializes state to its VM so it can pick up where it left off after a restart. Cursor's Aug 19 update ties Grok Bot into the /goal primitive shipped in the same week.

"Always-on digital coworkers that can handle real work across apps, inboxes, and tools." — Digital Applied's summary of the GA release

The Cursor bet

The tightest integration is with Cursor. Grok Bot inside a Cursor workspace can open, edit, and review code, run tests in the workspace VM, open PRs on Cursor Origin, and hand off code-review tasks to human reviewers. Cursor Pro+ users get one Grok Bot included; Ultra users get five; Teams get a per-seat allocation. The product is the first concrete deliverable of the SpaceX-Cursor acquisition that closed August 14, and it is the pattern xAI has been telegraphing: use Cursor as the primary developer distribution surface for xAI's frontier-model capabilities.

Pricing footprint

xAI's pricing has not moved with the GA release. SuperGrok Plus remains $30/mo, SuperGrok Heavy $300/mo, Cursor Pro+ $40/mo, Ultra $200/mo, Teams $50/seat/mo. Grok Bot compute is included in each plan up to a monthly VM-hour cap that xAI has not published; support forum posts suggest 100 hours per bot per month for Plus and unlimited for Heavy and Ultra. The Colossus 2 backend that powers Grok 4.6 is the same silicon fleet that services X.com Grok queries; Grok Bot workloads are being placed on separate priority queues to avoid degrading interactive Grok latency.

TechApplication

OpenAI ships Codex Remote GA, ChatGPT Voice in Codex desktop, and Sites URL swap

OpenAI Release Notes · Releasebot · 2026-08-21 · Source ↗

Friday's ChatGPT release notes promoted Codex Remote to general availability across every ChatGPT plan, added Voice mode inside the Codex desktop app, and let Site owners on Plus and Pro rename an existing Site's ChatGPT-hosted URL without redeploying. The update also expanded Codex local-projects to multi-folder workspaces and synced pinned chats between desktop and iOS.

OpenAI's Friday release notes tightened the developer surface across three surfaces at once. Codex Remote, the phone-to-desktop hand-off that lets users kick off a Codex job on a paired Mac or Windows host and monitor progress from an iOS or Android device, is now GA on every ChatGPT plan. Codex's desktop app added ChatGPT Voice, so users can speak an objective and hear the assistant think through it while inspecting file diffs on-screen. Multi-folder local-projects support let Codex hold several code roots in a single session with a designated primary folder for chats and Git operations. The pinned-chats sync between desktop and iOS covers all ChatGPT surfaces; a chat pinned on the Mac stays pinned on the phone within a few seconds.

ChatGPT Sites URL swap

Site owners on Plus and Pro can change an existing Site's ChatGPT-hosted URL without redeploying it. The previous address redirects to the new URL, preserving all routes and query parameters. The change closes the highest-friction pain point on the Sites product since it launched in June: brand rename or repositioning previously forced a full redeploy and dropped inbound links. OpenAI has not raised the per-Site pricing to accommodate the new capability.

"On Apple silicon Macs, the Apple Messages plugin in the ChatGPT desktop app can read and search iMessage, SMS, and RCS conversations and prepare or send messages through Messages." — OpenAI ChatGPT release notes, Aug 21

Codex Voice

Voice mode inside Codex is the first place OpenAI has shipped ChatGPT Voice inside a developer product. The implementation uses the same streaming voice pipeline that ships in the consumer app, so pause-and-resume, interruption, and multi-turn barge-in all work identically. Codex uses the voice input for objective setting and clarification, then hands control back to text-mode diffs and file edits for the actual code changes.

Enterprise governance

The Codex Remote GA rollout applies to individual accounts on all plans, but enterprise governance controls added earlier this month let admins gate Codex Remote at the tenant level. Enterprise Workspaces can also require that Codex jobs run on hosts inside a specific tenant-approved list, which is the compliance affordance that OpenAI put in place to sell into regulated industries where a laptop-to-server-anywhere pattern would not clear a security review.

ApplicationPolicy

Apple Music will label "Made With AI" songs, shifting responsibility to distributors

Billboard · MacDailyNews · Thurrott · 2026-08-21 · Source ↗

Apple emailed music industry partners Thursday spelling out a Made With AI labeling policy that will roll out to Apple Music listeners later in 2026. Distributors will be required to attach an AI Transparency Tag whenever generative AI is used to create a material portion of a track; enforcement mechanics and specific rollout dates were not disclosed.

Apple Music laid out a Made With AI labeling framework for its distributor network in a partner email dated Thursday, obtained by Billboard. Tracks that content providers identify as being materially generated using AI will begin displaying a Made With AI label on Apple Music service later in 2026. The framework requires distributors to attach an AI Transparency Tag inside the standard metadata payload whenever a material portion of a track is created by generative AI. Apple has not yet published exact launch date, threshold definitions for "material portion," or enforcement mechanics; the policy is being rolled out to distributors first so the metadata plumbing can be wired before the consumer-facing badge appears.

Why now

Apple Music VP Oliver Schusser told Billboard in April that more than a third of monthly uploads to Apple Music are 100 percent AI-generated, while AI music accounts for less than 0.5 percent of actual listening. The gap between upload volume and listening share is the operational problem Apple is trying to address. Making the label attach at the distributor level puts the compliance burden on Believe, DistroKid, TuneCore, CD Baby, and other aggregator distributors that already handle metadata contracts with Apple, rather than on Apple's own detection pipeline. Deezer earlier this year published detector-based figures showing that AI-only uploads have exceeded 45 percent of daily new music inflow across streaming services.

"Content providers will be required to include an AI Transparency Tag whenever AI is used to create a material portion of a track." — Billboard, summarizing Apple's partner email

What "material" means

Apple did not define materiality in the initial partner communication. The industry expectation, based on similar labeling rules that Deezer, Spotify, and YouTube Music have discussed, is that a track will trigger the tag when generative AI produces the vocal, the primary melody, or the arrangement, but not when AI is used only for mixing, mastering, stems separation, or vocal cleanup. Apple's framework leaves the initial classification with the distributor and reserves the right for Apple to audit and require corrections, with sanctions for repeated misclassification.

Roll-out sequence

Apple's Made With AI initiative extends the company's broader AI Transparency Tags framework announced earlier in 2026, which covers image and video content. The music tag is the second consumer-facing surface Apple has committed to labeling. No specific launch date was named beyond "later this year"; the industry expectation is a Q4 rollout timed to fall product releases.

ApplicationPolicy

Nevada clears up to 8,000 robotaxis for Tesla, Waymo and Uber in Clark County

TechCrunch · Rebecca Bellan · Engadget · 2026-08-21 · Source ↗

Nevada Transportation Authority commissioners unanimously approved three commercial robotaxi permits Thursday: Tesla can deploy up to 5,000 vehicles, Waymo up to 1,000, and Uber up to 1,000 through Motional and Zoox partnerships, across Clark County. Zoox retains a separate 100-vehicle permit granted earlier. Tesla Cybercab chief engineer Eric Early said 5,000 has always been a ceiling and the company would be satisfied hitting 2,500.

Nevada Transportation Authority commissioners unanimously approved three commercial robotaxi permits Thursday, clearing Tesla, Waymo, and Uber to operate driverless taxi services across Clark County over the next twelve months. Tesla's permit authorizes up to 5,000 vehicles, Waymo up to 1,000, and Uber up to 1,000 through its Motional and Zoox operator partnerships. Zoox retains a separate 100-vehicle permit granted earlier in the year. The combined 8,100 authorized vehicles is the largest single-jurisdiction robotaxi authorization on the books.

Tesla's ceiling versus its floor

Tesla asked for and received the 5,000 ceiling, up from an initial 10-vehicle cap the NTA imposed in late July that had drawn public criticism. Cybercab chief engineer Eric Early was frank about the near-term deployment reality: "The 5,000 has always been a ceiling for us…I think we would be extremely happy and satisfied if we could get ourselves up to 2,500." Tesla's Las Vegas service currently operates fewer than 100 vehicles, all inside a geofenced Strip corridor with a 45 mph top speed and no airport pickups. The Aug 20 permit expands the geographic scope to the entire state provided Tesla notifies the NTA of area expansions, but the vehicle count depends on production and safety-case ramp that Tesla has not yet demonstrated at scale.

Waymo and Uber

Waymo's 1,000-vehicle cap slots into a Vegas launch that is already several hundred vehicles ahead of Tesla in daily paid rides through the Waymo One app. The company won California PUC statewide authorization earlier this month for 18 counties covering the Bay Area, LA, Sacramento and San Diego, and the Nevada expansion completes a west-coast contiguous footprint. Uber's authorization runs through partnerships with Motional (downtown Las Vegas and Town Square) and Zoox (Strip corridor); Motional and Zoox handle the vehicles and Uber handles dispatch.

"Oversaturation of the commercial transportation industry and overcrowding of the roadways." — Kimberly Maxson-Rushton, representing the Livery Operators Association

Opposition and next steps

The Livery Operators Association and Vegas taxi companies opposed the permits at the Thursday hearing on the grounds cited above. The commissioners approved the permits despite the opposition after finding that the safety cases the three companies presented met the state standard. Uber advocated for a hybrid human-and-robotaxi approach in its filing. All three companies have already begun communicating rollout schedules to riders.

EconomicApplication

Twin1 launches with $20 million seed to build AI digital twins of knowledge workers

Axios · Ryan Lawler · TechStartups · 2026-08-21 · Source ↗

San Mateo-based Twin1 AI came out of stealth Friday with a $20 million seed round to build AI systems that represent individual professionals and can answer questions or take actions based on that person's knowledge, judgment, and work context. The four founders (Lewis Liu, Tom Cahn, Huiting Liu, Jonathan Budd) came out of document-AI company Eigen Technologies; Slack is the first channel integration.

Twin1 AI launched publicly Friday with $20 million in seed funding for an AI product that builds a digital twin of an individual knowledge worker. The San Mateo company was founded by four Eigen Technologies alumni — Lewis Liu, Tom Cahn, Huiting Liu, and Jonathan Budd — with the intent of packaging one professional's expertise into an on-call assistant that colleagues and clients can query directly. The company's initial target is professional services, specifically legal, where a partner's decisions on a complex matter can be encoded as reasoning traces the twin can replay against similar future matters.

How Twin1 is described

The product's own description is that it "answers questions and carries out tasks based on an individual's knowledge, judgment, and work context." The system is trained on a person's document history, email threads, meeting notes, and past work outputs, plus a structured layer of decision templates the person authors during onboarding. Twin1 exposes the twin through a Slack integration at launch, with API access on the roadmap. The value proposition is scaling one senior professional's judgment across an organization without requiring that person to be on every call.

Founder pedigree

Lewis Liu was CEO of Eigen Technologies, the London-based document intelligence company that raised more than $60 million from Goldman Sachs, Temasek, and Molten Ventures. Tom Cahn, Huiting Liu and Jonathan Budd worked with him at Eigen on the document-AI stack. The founding team's specific experience with contract intelligence and structured extraction sits directly upstream of what Twin1 is trying to do at the person layer: turn a professional's document history into a queryable knowledge substrate. The company has not disclosed the lead investor on the seed round in the Axios reporting.

"Answer questions and carry out tasks based on an individual's knowledge, judgment, and work context." — Twin1 product description

Where this fits in the market

Twin1's angle differs from Sana Labs' AI colleague product and from Rewind's personal-AI approach in that Twin1 is explicitly organizational: the twin is intended to be interrogated by other people, not to serve as an assistant to the person it represents. That positioning brings it into the same competitive lane as Harvey (legal AI) and Hebbia (professional services) but at a much narrower scope, and with a per-person pricing model rather than per-seat. Legal-services buyers who have watched partner burnout and knowledge-loss compound will be the early tell on whether the twin model finds product-market fit at senior-professional pricing.

Economic

Astromech raises $20 million at $3.8 billion valuation for biological forecasting

SiliconANGLE · The Next Web · Dealroom · 2026-08-21 · Source ↗

Dallas-based Astromech, the AI biology spinout of Ben Lamm's Colossal Biosciences, raised $20 million at a $3.8 billion valuation to develop foundation models that predict how biological systems change over time. Bob Nelsen led alongside Peak 6, NeoGenesis Capital, Builders VC, and CA Investments; total capital raised now $60 million after a $2 billion valuation just five months earlier.

Astromech raised $20 million at a $3.8 billion valuation, the Dallas-based AI biology startup disclosed Thursday. The company is co-founded by entrepreneur Ben Lamm and Harvard geneticist George Church, who together also built Colossal Biosciences, and its stated purpose is to build "a biological operating system that can forecast evolutionary change." The new round was led by Bob Nelsen alongside Peak 6, NeoGenesis Capital, Builders VC and CA Investments, bringing total capital raised to $60 million. The March round valued Astromech at $2 billion; the fresh $3.8 billion mark is nearly a doubling in five months.

What Astromech is building

Rather than train foundation models on language or images, Astromech combines genomic, evolutionary and functional biological data into a shared substrate that its models learn from. The stated goal is to identify how traits emerge, where biological systems become vulnerable, and how organisms, diseases and populations may change in the future. The commercial targets are drug discovery, aging research, and pandemic preparedness. The company remains in a research-heavy phase and has not yet published a paper on its underlying architecture; the current pitch is that the model class exists and is beginning to produce testable predictions inside partner wet labs.

"A biological operating system that can forecast evolutionary change." — Astromech's stated mission

The Colossal DNA

Astromech spun out of Colossal Biosciences, the woolly-mammoth and thylacine de-extinction company Lamm co-founded with Church. Colossal itself sits at roughly $10 billion in recent private-market marks. The spinout structure gives Astromech independent governance and a dedicated cap table while retaining shared research access to Church's Harvard lab and a portion of Colossal's compute and wet-lab infrastructure. Church is Astromech's chief scientist; Lamm is CEO. The round validates a specific bet that biology's foundation-model moment is now valuable enough to spin out as a standalone entity ahead of any published capability demonstration.

The counter-argument

Astromech is one of a growing class of biology-AI companies at multi-billion-dollar valuations before shipping a customer-facing product. Isomorphic Labs, Xaira Therapeutics, EvolutionaryScale, and Recursion sit at similar or larger marks with mixed public evidence. The bet from investors on the Astromech round is that Lamm's operating history and Church's scientific credibility together buy a longer runway than would be available to a first-time founder, and that the tissue between genomic data and predictive modeling is the specific piece the field is under-serving right now.

EconomicPolicy

Brazil unveils $444 million AI investment with US and Chinese partners

Reuters · TechStartups · 2026-08-21 · Source ↗

Brazil unveiled a $444 million package Friday to fund domestic AI capacity, with roughly $250 million going to a supercomputing project with Huawei and iFlytek. The rest funds Portuguese-language model training and strengthens applications for South American markets, positioning Brazil to work with both US and Chinese tech partners in the same national program.

Brazil's government unveiled a $444 million AI investment package Friday designed to build national AI capacity by partnering with both US and Chinese tech firms in a single program. The largest chunk, roughly $250 million, funds a domestic supercomputing project led by Chinese partners Huawei and iFlytek. The remaining $194 million supports Portuguese-language model training, applications for South American markets, and public-sector AI adoption across federal ministries. The package extends the AI Plan for Brazil that President Lula introduced in 2024, which had targeted 23 billion reais of AI investment through 2028.

"Demonstrating a multipolar AI infrastructure race." — Reuters, on Brazil's dual-partner approach

The Huawei-iFlytek arrangement

The supercomputing project puts Huawei's Ascend silicon and iFlytek's LLM stack at the base of a national compute facility Brazil can offer to domestic labs and universities. The specific configuration and site have not been named. iFlytek has been building a Portuguese-language variant of its Spark model at the request of Brazilian government partners over the past two years, and the Friday announcement provides the compute host that variant will run on. Brazil's decision to work with a US-sanctioned Chinese chip supplier for the compute layer is explicit rather than accidental; the country's foreign minister has publicly argued for non-alignment on AI infrastructure specifically to avoid being caught between US and Chinese export-control regimes.

The US component

Portions of the $194 million balance flow to research collaborations with US-based labs and to US-linked cloud infrastructure that Brazil intends to hold in domestic data centers. Names on the US side were not disclosed in the Reuters reporting. Brazil's stance on AI has been that the country buys from whoever will sell without conditions that would force Brazil to choose sides, and Friday's package is the operational expression of that policy.

Portuguese-language AI

The Portuguese-language model line is the piece of the investment that has near-term downstream consequences for Brazil's public-sector AI adoption. Federal agencies have been running pilots on Claude, GPT-5.6 Sol, Qwen 3.8 and Gemini 3.6 Flash, none of which were specifically trained on Brazilian Portuguese at frontier capability. Building a domestic model line explicitly on Brazilian Portuguese reduces the linguistic-quality penalty government agencies have been eating on English-trained models applied to national workloads.

TrendEconomic

New York passes San Francisco as top US tech talent market for the first time

CNBC · CBRE · Quartz · Commercial Observer · 2026-08-21 · Source ↗

CBRE's annual Scoring Tech Talent report puts New York at 394,300 tech workers against San Francisco's 375,730 for the first time in the report's 13-year history, per data published Friday. Finance and AI hiring drove the New York flip: 90,530 tech jobs added in the finance/insurance/real estate sector since 2022 while the high-tech industry cut 21,262 in the same window.

New York's tech talent headcount surpassed San Francisco's for the first time in CBRE's Scoring Tech Talent series, per the annual report released Friday. The city recorded 394,300 tech workers against San Francisco's 375,730, ending a 13-year San Francisco lead. The flip is being driven by two opposing forces: layoffs of non-AI tech workers in the Bay Area concentrated in 2023-2025, and strong AI-plus-finance hiring in New York over the same window. Since 2022, the finance, insurance and real estate sector added 90,530 tech jobs across the US while the high-tech industry cut 21,262 in the same period.

What is behind the flip

New York's tech growth is not concentrated in AI companies specifically. It is concentrated in AI hiring at banks, insurers, and law firms that until recently would have staffed those roles as generalist software engineers or quantitative analysts. Goldman Sachs, JPMorgan, Bank of America, Bloomberg, and the AmLaw 20 firms have been rebuilding entire teams around ML infrastructure and applied AI over the past 24 months. That shift maps directly onto New York's headline count and explains why the city's tech-talent share is up even as venture-funded AI companies still cluster in the Bay Area.

"San Francisco retains its position at the top of CBRE's indexed rankings, which weight 13 metrics including tech talent concentration, talent pipeline, and research and development investment. San Francisco scored 81.98 in that index, compared with New York Metro's 70.38." — CBRE's Scoring Tech Talent 2026 report

SF still wins the composite

Headcount is only one input into CBRE's composite score, which weights 13 factors including talent pipeline depth, R&D investment, and concentration. San Francisco tops the composite at 81.98 versus New York Metro's 70.38. The gap says that the Bay Area still has proportionally more AI research depth and R&D concentration; New York's headcount lead is driven by absolute volume of finance-side AI hiring rather than by AI-native company density.

Where AI hiring grew fastest

Both cities added more than 20,000 AI-specific jobs since mid-2025, per the CBRE data. AI-skilled worker headcount grew 45 percent year-over-year across US and Canadian tech markets. Third-place Toronto, fourth-place Seattle, and fifth-place Boston are all showing above-market AI growth as well, but the New York and San Francisco flip is the specific headline outcome of the current cycle.

Policy

Judge tosses economic-espionage counts against ex-Google engineer Linwei Ding

Reuters · Investing.com · BusinessWorld · Benzinga · 2026-08-21 · Source ↗

US District Judge Vince Chhabria in San Francisco threw out the seven economic-espionage convictions against former Google engineer Linwei Ding on Thursday, ruling there was insufficient evidence Ding knew his conduct would benefit the Chinese government. The seven trade-secret theft convictions remain; sentencing is scheduled for September 1. Ding was found guilty in January of stealing thousands of pages of confidential AI infrastructure details.

Judge Vince Chhabria tossed the seven economic-espionage counts against former Google software engineer Linwei Ding on Thursday, ruling that federal prosecutors did not present sufficient evidence that Ding intended or knew his conduct would benefit the Chinese government. The seven separate trade-secret theft convictions Ding was found guilty of in January stand. Ding is scheduled to be sentenced on September 1 in the Northern District of California. The ruling is a partial win for the defense and cuts materially into the total prison exposure Ding faces at sentencing, without disturbing the underlying finding that he stole Google's AI infrastructure documents.

What the government proved and what it didn't

Prosecutors persuaded the jury in January that Ding stole thousands of pages of confidential Google documents describing the hardware infrastructure and software platform that trains the company's large AI models. That underlying theft finding is untouched by Thursday's ruling. What Chhabria rejected is the further government contention that the theft was for the benefit of the Chinese state. The economic-espionage statute requires proof of intent or knowledge that the target of the theft benefits a foreign government. Chhabria wrote that the trial record contained circumstantial evidence about Chinese-headquartered companies and Ding's contacts, but did not clear the bar for the intent element on the espionage counts. The trade-secret theft counts, which require only proof of theft plus commercial value, remain intact.

"Insufficient evidence that Ding intended or knew his conduct would benefit the government of China." — Judge Vince Chhabria

What Ding took

The documents at issue describe Google's TPU pod architecture, the software stack that partitions large-model training runs across TPU pods, and internal specifications for the datacenter-scale infrastructure that Google uses to train Gemini. Ding was a Google software engineer with legitimate access to those documents and used that access to move copies onto personal accounts while accepting positions with two China-headquartered AI startups. He was arrested in March 2024 and stood trial in late 2025.

What sentencing looks like now

Without the economic-espionage counts, Ding's sentencing exposure is bounded by the trade-secret theft convictions and the associated federal sentencing guidelines. Prosecutors are expected to press for a substantial custodial sentence on the trade-secret side even without the espionage counts. Defense counsel is expected to argue for a lesser sentence given the partial reversal. The September 1 date sits ten days out from Friday's ruling.

PolicyEconomic

Supermicro board investigation clears senior management in $2.5 billion Nvidia chip diversion

Fortune · The Register · Bloomberg · Tom's Hardware · 2026-08-21 · Source ↗

Super Micro Computer's board disclosed Thursday that its independent investigation into the $2.5 billion Nvidia-chip diversion scheme found no evidence current senior management knew about the alleged smuggling. The company terminated several employees; co-founder Yih-Shyan "Wally" Liaw's criminal trial was pushed from November 2026 to March 2027.

Super Micro Computer's board of directors said Thursday that an independent investigation, conducted by an outside law firm with a forensic-accounting consultant, found no evidence that current members of senior management knew about the alleged scheme to smuggle $2.5 billion of hardware packed with Nvidia chips to China. The company disclosed that several employees were terminated in connection with the findings but did not name them. The investigation was launched after the March indictment of Supermicro co-founder Yih-Shyan "Wally" Liaw, along with Taiwan-based executive Ruei-Tsang "Steven" Chang and alleged broker Ting-Wei "Willy" Sun, on charges of conspiring to violate the Export Controls Reform Act.

"Neither the company nor its current senior executives were part of the alleged AI chip smuggling." — Super Micro Computer, on the independent investigation findings

How the scheme was described

The Justice Department's March indictment described a multi-year operation in which Supermicro servers loaded with Nvidia H100 and H200 GPUs were diverted through re-export channels to buyers in mainland China, in violation of the US export-control regime that requires end-use certifications for shipments to Chinese customers. The alleged conspiracy period spans 2023 and 2024 and the $2.5 billion total is the government's estimate of end-user hardware value that reached prohibited destinations. Liaw has pleaded not guilty; his trial was pushed back from November 2026 to March 2027 after the defense sought more time on discovery.

What the internal report changes for Supermicro

The board's finding provides the paper trail Supermicro needs to defend the company's continued eligibility for US federal contracts and for continued Nvidia allocations. Supermicro shares had cratered 33 percent on the March indictments; the internal exoneration on senior-management involvement was a specific piece the market had been waiting for. Shares recovered on Thursday's disclosure. The company disclosed the termination of several employees but did not disclose whether the terminations covered the individuals named in the indictment or a wider group implicated by the investigation.

The bigger export-control frame

The Supermicro case is the largest documented AI-chip export-control breach in the current export-control regime, and its resolution sits at an inflection for US-Chinese semiconductor policy. If the March 2027 trial concludes without top-management liability sticking to Supermicro's current officers, the case cements a pattern where individual executives absorb export-control criminal exposure while corporate parent liability stops at internal-controls remediation. The Justice Department has not commented on the internal-investigation findings.

Tech

Hugging Face Papers of the Day cluster around agents and 4D generation

Hugging Face · arXiv · 2026-08-21 · Source ↗

Friday's Hugging Face Papers of the Day led with Ant Research's 4DAnyone at 162 upvotes, followed by Google DeepMind's EnvHarness at 54, OpenMOSS SWE-bench Science at 44, and Tencent Hunyuan WithEveryone at 38. USTC's FACET, ZJUNLP's MemTrapBench round out the top set.

Hugging Face's Papers of the Day for August 21 clustered around embodied agents and generative video/3D work. Ant Research's 4DAnyone led at 162 upvotes; Google DeepMind's EnvHarness came in second at 54; OpenMOSS's SWE-bench Science at 44; Tencent Hunyuan's WithEveryone at 38; USTC's FACET at 23; and ZJUNLP's MemTrapBench at 1.

Ant Research 4DAnyone

4DAnyone (arXiv:2608.20335) proposes reconstructing high-fidelity 4D humans from single-camera monocular footage by first generating multiview-consistent videos and then lifting them into 4D Gaussian Splatting. Two key contributions: Reference Context Packing compresses reference views into fixed-length context, and Target Context Routing dynamically manages view groupings during the denoising process to maintain structural consistency across time. Authors: Yudong Jin, Tao Xie, Qihang Zhang, Zehong Shen, Zhen Xu, Yujun Shen, Hujun Bao, Xiaowei Zhou, Yinghao Xu. The work sits inside Ant's Human Body Reconstruction program that produced multiple SIGGRAPH 2025 papers on avatar reconstruction.

"Reference Context Packing compresses reference views into fixed-length context." — 4DAnyone, on the scaling mechanism

Google EnvHarness and OpenMOSS SWE-bench Science

EnvHarness (arXiv:2608.19880) from Google DeepMind proposes turning static environments into interactive agent-training substrates by wrapping the static content in a lightweight simulator that responds to agent actions. The framework's motivation is that static worlds (websites captured to disk, screenshots of dashboards, offline app bundles) cannot be used to train agents that require environmental response, and the wrapper turns them into usable training data. OpenMOSS's SWE-bench Science (arXiv:2608.19799) extends the popular SWE-bench coding-agent benchmark with tasks drawn from computational-science engineering problems, testing whether coding agents that clear SWE-bench can also resolve engineering tasks embedded in scientific software repositories. Tencent Hunyuan's WithEveryone (arXiv:2608.20336) offers unified planning and identity grounding for group-image generation, holding N distinct subject identities consistent across scenes.

The rest of the top set

USTC's FACET (arXiv:2608.18580) preserves source intent and executable state during automated synthesis of terminal-based tasks, a pipeline aimed at building larger and more faithful terminal-agent benchmarks. ZJUNLP's MemTrapBench (arXiv:2608.20202) is a benchmark specifically for cognitive traps in LLM memory use, measuring how frequently language models fall for adversarial memory patterns that induce hallucinated retrieval. All papers include public code releases per Hugging Face's paper-page policy.

SafetyTech

Patch Thursday: Zimbra, Cisco Crosswork, and Head Mare's PhantomCore across 22 disclosures

eSecurity Planet · CERT Polska · Cisco PSIRT · 2026-08-21 · Source ↗

Twenty-two patches shipped Friday covering code execution, privilege escalation and information disclosure. CERT Polska confirmed active exploitation of Zimbra Collaboration CVE-2026-73570; four Cisco Crosswork disclosures cover Data Gateway, Network Controller and Planning regardless of device configuration; Head Mare hacktivists are chaining several of the bugs to deploy PhantomCore malware.

Twenty-two security patches shipped Friday across enterprise infrastructure vendors, per the eSecurity Planet week-in-security roundup. The bulk of the fixes cover code-execution, privilege-escalation, and information-disclosure vulnerabilities. Three separate operational stories sit under the disclosure cluster: CERT Polska confirmed active exploitation of a Zimbra Collaboration vulnerability tracked as CVE-2026-73570; Cisco published four Crosswork advisories covering Data Gateway, Network Controller and Planning products; and the Head Mare hacktivist group is chaining several of the disclosed bugs to deploy the PhantomCore malware family in operational campaigns.

Zimbra CVE-2026-73570

The Zimbra Collaboration vulnerability is under active exploitation, per a CERT Polska advisory Friday. The vulnerability sits in a pre-authentication code path and allows an attacker to execute arbitrary code on a vulnerable Zimbra server. Zimbra Collaboration is deployed at governments, universities, and enterprises across Europe as a webmail-plus-collaboration stack, and CERT Polska's advisory framed the exploitation as opportunistic rather than targeted. Administrators were told to apply the Zimbra patch immediately and to inspect servers for the specific indicators of compromise the advisory publishes.

Cisco Crosswork disclosures

Cisco's PSIRT published four separate Crosswork advisories covering three products: Crosswork Data Gateway, Crosswork Network Controller, and Crosswork Planning. The advisories affect deployments regardless of the specific device configuration, which is the specific phrasing Cisco uses to say the vulnerable code path is exercised by default without unusual configuration required. Crosswork is Cisco's automation and orchestration platform for service-provider networks; the advisories require patch application because service-provider workloads sit downstream of Crosswork's automation planes.

"Exploitation of the Zimbra Collaboration vulnerability CVE-2026-73570 has been observed by Poland's CERT Polska." — eSecurity Planet, Aug 21 weekly roundup

Head Mare and PhantomCore

The Head Mare hacktivist group has been observed weaponizing several of the Aug 21 disclosures to deploy PhantomCore malware. PhantomCore is a modular loader that has been in circulation since 2024 with successive variants adding privilege-escalation and credential-harvesting modules. Head Mare's operational focus has been Russian and Belarusian government and defense targets in support of Ukraine, and the group's rapid weaponization of newly-disclosed bugs suggests either advance knowledge of the patch cycle or an efficient patch-diffing pipeline of its own. Defenders on the Ukrainian side of the front are the primary beneficiaries of the exposure; enterprise IT teams outside the region are advised to patch on standard schedules and treat PhantomCore signatures as commodity.

Tech

llama.cpp cuts build b10549 at 09:23 UTC with OpenVINO 2026.3 across ARM64 and CUDA 13.4

GitHub ggml-org · 2026-08-21 · Source ↗

The ggml-org llama.cpp project cut build b10549 at 09:23 UTC Friday, publishing binaries across Windows CUDA 12.4 and 13.3/13.4, Android ARM64, macOS ARM64 and x64, and Ubuntu ARM64 with the OpenVINO backend bumped to 2026.3. The release completes the OpenVINO version alignment that started earlier in the week with b10502.

llama.cpp shipped build b10549 at 09:23 UTC on Friday, with additional per-platform binaries stamping through the following hours. The release publishes binaries across Windows CUDA 12.4 and CUDA 13.3/13.4, Android ARM64, macOS ARM64 and x64, and Ubuntu ARM64 with the OpenVINO backend on 2026.3. The OpenVINO version bump completes the alignment that started with b10502 earlier in the week; downstream forks that had been carrying local patches for OpenVINO 2026.3 support can now target the upstream build directly.

What the release covers

The build is a routine cut with the primary payload the OpenVINO version bump plus the standard set of quantization performance improvements, tokenizer fixes, and platform-specific compile flags that llama.cpp merges through each release. The signed-attestations infrastructure introduced in b10502 continues to cover this build, so consumers verifying binary provenance through GitHub's attestation mechanism can rely on the same tooling. The maintainers have been on a Friday-cut cadence for several weeks; b10549 is the second Friday release of the month and continues the pattern.

Why OpenVINO 2026.3 matters

The OpenVINO 2026.3 alignment matters for Intel Arc B-series GPU users, who saw the quantized-copy fix in b10502 lift q4_0-to-f32 throughput 7.8x from 20.21 to 158.19 GB/s on Arc 70 hardware. The b10549 build carries that improvement forward and paves the way for the Intel Arc B60 kernel work that has been open in draft PRs for two weeks. Intel and llama.cpp maintainers have been closer in coordination than at any previous point in the project's history, and the OpenVINO cadence reflects that alignment.

"OpenVINO backend bumped to 2026.3 across ARM64 and CUDA 13.4 platforms." — Build note for b10549

Consumer implications

For end users, the upgrade from b10502 or earlier is safe and recommended for anyone running Intel Arc GPUs or targeting the OpenVINO backend on modern CPUs. Users on CUDA 12.4 can stay on their current install without functional loss; the CUDA 13.4 build is required only for Blackwell RTX 5000-series GPUs. macOS ARM64 build users on M3 or M4 silicon should upgrade to pick up the tokenizer edge-case fix that landed alongside the release notes.