IMPLICATOR
San Francisco | Tuesday, June 2, 2026
Anthropic has filed the quietest IPO paperwork around the largest startup valuation in AI. The company says little in public. The market sees $965 billion, $47 billion of annualized revenue and an S-1 that could put its SpaceX compute bill into the open.
Nvidia is pulling the hardware side closer to the desk. RTX Spark puts CUDA and up to 128GB of unified memory into Windows PCs, while Nemotron 3 Ultra gives US enterprises a fast open model that still trails China's Kimi K2.6 on the same index.
Both stories end with receipts. Anthropic eventually has to show margins, obligations and charity pledges. Nvidia has to show prices, thermals and buyer demand when the fall systems ship.
Stay curious,
Marcus Schuler
Know someone drowning in AI noise? Forward this briefing. They can subscribe free here.
Anthropic Confidentially Files for IPO at $965 Billion, Beating OpenAI
Anthropic submitted a confidential draft S-1 to the SEC on Monday, putting it ahead of OpenAI in the race to take a frontier AI lab public. The filing follows last week's $65 billion funding round that valued the company at $965 billion. Revenue hit $47 billion on an annualized run rate, though those figures are self-reported and unaudited.
The filing starts a regulatory clock without disclosing financials. A confidential draft lets Anthropic begin the SEC review while keeping private until the fall the numbers that would determine whether its lead over OpenAI is durable. Anthropic said the offering "will depend on market conditions and other factors." SpaceX goes public first, on June 12, at a valuation near $1.75 trillion. OpenAI is targeting September.
The comparison between the two rivals is uneven because the numbers point to different moments. Anthropic's $965 billion valuation came from last week. OpenAI's $852 billion dates to a $122 billion round in March. Anthropic's $47 billion run-rate is unaudited, and the Wall Street Journal reported unclear accounting. A SpaceX compute deal already commits roughly $1.25 billion a month through May 2029, close to a full year of Anthropic's claimed revenue.
Why This Matters:
- Public market investors are about to price three trillion-dollar AI IPOs in a single year, and the self-reported numbers behind the first filing will set the baseline. Anthropic's S-1 will surface SpaceX, Amazon, Google and Broadcom obligations that currently sit off the public record.
- The fall calendar now holds a test that filing first cannot pass. A public S-1 will put compute costs, margins and the funding math behind that $47 billion figure on the record.
Reality Check
What's confirmed: Anthropic confidentially filed a draft S-1 on Monday at a $965 billion valuation with $47 billion in claimed annualized revenue.
What's implied (not proven): The self-reported revenue uses accounting methods that make the run-rate comparable to public-company revenue, and the SpaceX compute deal is the largest single obligation.
What could go wrong: A public S-1 reveals margins, obligations or accounting choices that surface a lower effective valuation once public-market analysts apply standard metrics.
What to watch next: The public S-1 this fall, the first moment anyone outside the company can verify the numbers.
The One Number
2 seconds - Tripo AI says its P1.0 model can generate production-ready 3D meshes in as little as two seconds. The claim matters because manual asset work still runs through cleanup, texture and export checks before a game or studio pipeline accepts it. The funding bet is that AI 3D moves from demo object to usable asset.
Source: The Implicator, June 1, 2026
💰 Fresh Funding
💰 Fresh Funding
Raises nearly $200M: Tripo AI expands AI 3D and world-model roadmap
Tripo AI said Monday it completed Series A+ and A++ financings totaling nearly $200 million, with Ince Capital and a China Life-backed Yangtze River Delta tech fund leading according to Implicator coverage and company statements. The company is pushing its Tripo 3D platform toward production meshes and reusable AI-generated worlds for game, film and design teams.
Visit Tripo AI →Discloses $60M: Mecka AI trains robots with human motion data
Fortune reported Monday that Mecka AI disclosed $60 million across a $25 million Series A and a $35 million follow-on investment led by Framework Ventures. The company captures human movement through body sensors and iPhones to build training data for robotics teams that need real-world actions alongside simulated motion.
Visit Mecka AI →Raises €20M: Invisix brings soft X-ray metrology to AI-era chips
Business Wire said Monday that Invisix raised an oversubscribed €20 million seed round from Hitachi Ventures, Transition Ventures, imec.xpand, Doosan Investment and a tier-one chipmaker. The ASML spinout uses soft X-ray metrology to inspect buried chip structures as AI processors become harder to measure without destroying wafers.
Visit Invisix →Nvidia Pushes RTX Spark Into Windows PC Processor Slot This Fall
Nvidia said at Computex that RTX Spark will put a Blackwell-class GPU, 20-core Grace CPU and up to 128GB of unified memory into Windows laptops this fall, moving the company beyond add-in graphics into the processor slot held by Intel, AMD and Qualcomm. More than 30 laptop designs and 10 compact desktops are planned from unnamed partners.
That hardware pitch carries a software wrapper Nvidia has not bundled before. OpenShell, a secure agent runtime, brings local AI agents to Windows with Microsoft's security primitives. Nvidia says the top configuration can run 120-billion-parameter models with a one-million-token context and render 90GB 3D scenes on 128GB of unified memory.
Battery life, Windows on Arm compatibility and pricing stay unpublished, and a company that chose not to publish benchmark charts made the "most efficient PC chip ever built" claim. Nvidia lists more than 100 software and game partners, including Adobe, Riot Games and Krafton, an anti-cheat compatibility test more than a graphics one.
AI Image of the Day
Prompt: a white cat on the head --ar 2:3 --raw --sref 2508009867 7909 --profile hjf34yn --hd --v 8.1
Nvidia Nemotron 3 Ultra Leads US Open Models, Trails China's Best
Nvidia unveiled Nemotron 3 Ultra at GTC Taipei on Monday, a 550-billion-parameter open-weight model that scores 48 on the Artificial Analysis Intelligence Index, the top US open model but six points behind China's Kimi K2.6. The model ships with a free enterprise Agent Toolkit designed to make Nvidia silicon the default hardware for running agents.
The toolkit includes NemoClaw for agent orchestration, OpenShell for secure runtime, and CUDA-X skills exposed to agents under open licenses. Cadence, CrowdStrike, Palantir, Siemens and Foxconn signed on as launch partners. Cadence built ChipStack, a chip-verification agent that Nvidia is using to verify its own silicon.
At 300 tokens per second on a pre-release endpoint at DeepInfra, the model runs up to five times faster than comparably sized rivals, according to Nvidia's own testing. The model is free, the toolkit is open source, and the speed that sells it runs fastest on the chips Nvidia sells.
🧰 AI Toolbox
How to Turn a One-Sentence Brief Into a Finished PowerPoint Deck with Riffly
Riffly is an AI deck generator that takes a short description of a presentation and builds a full slide deck you can export to PowerPoint, Google Slides, or PDF. Pick a tone (executive, pitch, lecture, training), feed it your topic plus any reference docs, and Riffly drafts the structure, writes the bullets, picks images, and applies a consistent design. Useful when you have 20 minutes to prep a presentation and need a real draft to edit, not a blank slide. Free tier covers short decks.
Tutorial:
- Go to tryriffly.app and sign in with Google or email
- Describe the deck in plain English: "10-slide Q2 board update covering revenue, product launches, hiring, and risks"
- Upload any reference docs (last quarter's update, current OKRs, a metrics dashboard export) so Riffly grounds the bullets in your real data
- Pick a template that matches your audience: clean executive, pitch deck, training material, or conference talk
- Let Riffly draft the full deck, then click any slide to edit text, swap images, or regenerate the layout
- Use the "Speaker notes" toggle to auto-generate notes you can read while presenting
- Export to PowerPoint, Google Slides, or PDF. The file keeps the design system so you can keep editing in your tool of choice
URL: tryriffly.app
What To Watch Next
|
||
|
||
|
||
|
||
|
💡 5-Minute Skill
Turn an AI Agent Pilot Into a Kill-Switch Plan
Tuesday, 8:37 a.m. Product wants an agent to reopen tickets, update Salesforce and issue small refunds. Security wants a hard no. The better answer is a pilot that can stop itself before a mistake becomes a customer incident.
Your raw input:
Agent: support-renewal assistant. Tools requested: Zendesk write access, Stripe refunds up to $250, Salesforce opportunity updates and Slack DMs to account owners. Pilot: 20 accounts for two weeks. Known risks: duplicate refunds, wrong account notes, private health data in tickets and ignored approval thresholds. Need: go/no-go checklist before legal review.
The prompt:
Act like a pragmatic AI security reviewer. Convert this agent request into a go/no-go checklist for a two-week pilot. Separate read-only access, supervised write access, refund authority, logging, human approval thresholds and rollback. Give me three kill conditions that end the pilot immediately. Use only the facts above.
The output:
Start read-only for Zendesk and Salesforce across the 20 pilot accounts. Move to supervised writes only after every proposed ticket change includes account name, source ticket and confidence reason. Refunds stay blocked until Finance approves a $250 cap and every refund requires a human click. Log tool calls, approvals, denials and rollback time. Kill the pilot after any duplicate refund, any write to a non-pilot account or any private health data copied into Slack.
Why this works:
Agent pilots fail when a small permission quietly becomes production authority. This prompt forces staged access, visible logs and stop conditions before the team argues about speed. The kill conditions matter because they turn risk into a rule someone can enforce on a bad Tuesday.
What to use:
Claude is best when you paste policy, access notes and messy ticket examples. ChatGPT is fine for a short checklist. Keep the phrase "kill conditions" in the prompt. Otherwise the model writes a safety plan that sounds responsible and stops nothing.
📖 AI Alphabet
|
S
|
📖 AI Alphabet Semantic Search Semantic search tries to find results based on meaning, not just exact keyword matches. That is why it can surface relevant material even when the query uses different words. |
AI & Tech News
Former Meta CTO Closes $250M Climate Fund for AI Infrastructure
Bloomberg reported that Mike Schroepfer's Gigascale Capital closed a $250 million fund for early climate startups tied to AI infrastructure. The pitch links cleaner power, cooling and efficiency to a compute buildout that is pulling more capital into grid-adjacent companies.
Salesforce's Anthropic Stake Reaches $5 Billion
Salesforce's early Anthropic investment is now valued at about $5 billion, according to Bloomberg, after an initial $50 million commitment in 2023 and follow-on rounds. The gain gives Salesforce a financial stake in the same Claude relationship it sells to enterprise customers.
GitHub Copilot Pricing Switch Angers Developers
Ars Technica reported that GitHub Copilot's new usage-based pricing started Monday and immediately produced user complaints about fast-depleting monthly caps. The reaction shows how AI coding tools are moving from flat-rate subscriptions to metered usage as developers lean on them for longer sessions.
Hackers Used Meta AI Support to Hijack Instagram Accounts
404 Media reported that attackers asked Meta's AI support system to change emails on high-profile Instagram accounts, and the system complied. Meta said it patched the flaw, but the episode gives every support chatbot buyer a hard security question: who can authorize account recovery?
Alphabet Seeks $80 Billion in Equity for AI Spending
Bloomberg reported that Alphabet plans to raise $80 billion in equity, with Berkshire Hathaway set to invest $10 billion, to fund AI infrastructure and product work. The size turns Google's AI budget into a capital markets event.
Salesforce Buys Contentful at Discount to 2021 Valuation
Techmeme linked to The Information's report that Salesforce acquired Contentful for about $1 billion to $1.5 billion in cash. The price sits below Contentful's reported $3 billion private valuation in 2021, a sign that AI budgets have not rescued every software exit.
Pre-ChatGPT Unicorns Face Private Market Reset
CNBC reported that more than half of US unicorns have not raised new capital in three years, with more than 220 now classed as fallen unicorns after valuation cuts. Investor appetite has moved toward AI-native companies, leaving older private software names with stale prices and tougher financing conversations.
HPE Shares Jump After Server Demand Lifts Outlook
Bloomberg reported that HPE shares surged after the company raised its growth forecast, citing strong server demand over the next 18 months. Revenue rose 40% year over year to $10.7 billion in the quarter, giving investors another read on AI infrastructure spending outside the hyperscalers.
Red Hat Cloud Services Packages Carry Credential-Stealing Malware
Step Security researchers found that multiple official @redhat-cloud-services npm packages were compromised with malware that stole tokens from GitHub Actions, AWS, Google Cloud and Azure environments. The code ran through a preinstall hook, which means developer machines could leak credentials before anyone opened the package.
Endra Raises $50M to Automate Building System Design
Axios reported that Stockholm startup Endra raised a $50 million round led by Andreessen Horowitz for AI software that designs plumbing and electrical systems in new buildings. The use case is less glamorous than humanoid robots, which may be the point: construction design has repetitive engineering work and expensive delays.
🚀 AI Profiles: The Companies Defining Tomorrow
Do Anything
Do Anything launched the alpha of a platform where AI agents run autonomously for months at a time, each with its own name, email address, and persistent identity. The pitch is that the next generation of agents will not respond to one-off prompts but will live on the internet as colleagues you assign work to and check in on. 🤖
Founders
Do Anything was launched out of an early-stage team building on the MEESEEKS framework, which orchestrates long-running agent workflows across tools and time. The founding team has not been publicly detailed beyond the launch coverage.
Product
Each agent on Do Anything's platform gets persistent memory, an email address that other humans and agents can reach, scheduled triggers, and access to a tool catalog (web browsing, code execution, API calls, communications). The product is designed so users can hand an agent a long-horizon goal ("monitor these competitors and email me weekly with material changes") and trust the agent to operate independently for weeks.
Competition
The autonomous-agent category includes Lindy, Hermes Agent, Skygen, MultiOn, and the agent frameworks from frontier labs (OpenAI's Agents SDK, Anthropic's Computer Use, Google's Agent Builder). Do Anything's wedge is the explicit "long-running colleague" framing, which is a UX bet as much as an engineering one.
Financing 💰
Do Anything launched its alpha in early 2026 with funding details not publicly disclosed at the time.
Future ⭐⭐⭐
Long-running agents with persistent identity are the most interesting unsolved UX problem in AI. The technical bar is high (agents that drift quietly, run up bills, or send confused emails will fail fast), but if Do Anything cracks the reliability problem, the product reframes how teams think about hiring versus deploying. 🛠️
🤨 Yeah, But...
The Implicator reported Sunday that Vast, the company behind Tripo AI, raised nearly $200 million and crossed a $1 billion valuation. Tripo says P1.0 can generate production-ready meshes in as little as two seconds, while Project Eden is meant to create reusable AI-generated worlds. (The Implicator, June 1, 2026)
Our take: Tripo's pitch has a wonderfully dangerous unit: seconds. Once a model claims it can make a production mesh before a producer finishes asking for one, every manual handoff starts looking like a tax. Studios know the catch. A game asset earns its keep after rigging, texture, licensing, style direction and the person who says the elbow looks cursed. The funding says investors think those chores shrink faster than artists can invoice for them. The first buyer of AI 3D is likely buying fewer blank hours between idea and review, with all the arguing about taste still included.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
