IMPLICATOR
San Francisco | April 14, 2026
OpenAI absorbs a personal-finance startup and the team that built it. Hiro Finance goes dark next week. Ethan Bloch, on his fifteenth fintech attempt, now builds ChatGPT's answer to Mint and the robo-advisors, the stickiest category in consumer software.
Meanwhile, CRO Denise Dresser accuses Anthropic of padding its $30 billion run rate by $8 billion. Gross versus net. Accounting trivia, until you remember both companies plan to file S-1s this year and Ramp data already shows Anthropic closing the enterprise gap.
And a 20-year-old from Texas arrives at Altman's Russian Hill gate at 3:45 a.m. Friday with a Molotov cocktail and a backpack list of other AI executives by home address. Prosecutors charged him with attempted murder on Monday.
Stay curious,
Marcus Schuler
Know someone drowning in AI noise? Forward this briefing. They can subscribe free here.
OpenAI Buys Hiro Finance in Acquihire, Puts ChatGPT on Collision Course With Mint
OpenAI quietly absorbed personal-finance startup Hiro Finance in an acquihire Monday. Founder Ethan Bloch and his whole team are moving to ChatGPT. Hiro stops taking signups immediately, shuts down April 20, and wipes all user data by May 13.
This is a people deal, not a technology deal. The servers are getting erased, the brand is going away, and nothing Hiro-shaped is showing up inside ChatGPT next week. What OpenAI bought is Bloch, co-founder Rushabh Doshi, and whatever verification work the team did to keep large language models honest on financial math, a problem frontier models have not yet fully solved.
Bloch is not a first-time seller. Digit, his algorithmic savings app, sold to Oportun in 2021 for around $230 million. Hiro was his fifteenth project. Across three fintech attempts, he has said the same thing about the same goal: build "a Jane," the AI assistant from Ender's Game who handles the main character's taxes. ChatGPT is now his third try at the same idea.
For OpenAI, the deal fits the recent pattern of killing side projects (Sora, the adult-content mode) and launching vertical ones (the $100 Pro tier around Codex). ChatGPT is becoming a set of products sitting on top of one interface. Personal finance is the newest vertical, and the one where no one has a financial services license yet.
Why This Matters:
- Personal finance is the stickiest relationship in consumer software. Whoever owns the default interface for net worth, debt, and retirement owns a moat that Mint, Betterment, and traditional wealth managers thought compliance had protected.
- Regulatory questions that were already coming just moved to this quarter. OpenAI has said nothing about whether it will seek a state financial advisor registration or how it will store the kind of data ChatGPT is about to start handling.
Reality Check
What's confirmed: OpenAI acquired Hiro in an acquihire, the whole team is moving, Hiro shuts down April 20, user data wiped May 13. Bloch's fifteenth project.
What's implied (not proven): ChatGPT is about to ship a dedicated personal-finance product in the next quarter or two.
What could go wrong: One hallucinated decimal in a retirement projection and the first state regulator piles in.
What to watch next: Whether OpenAI files for, acquires, or partners into a financial advisor registration inside Q2.
The One Number
2.8 gigawatts — Oracle's new fuel-cell capacity ceiling with Bloom Energy, announced four business days after Oracle received a $400 million stock warrant in the supplier. That is 40% more than Bloom's entire annual manufacturing capacity of 2 gigawatts. Oracle is not buying power. It is buying the factory.
Source: Implicator.ai
OpenAI CRO Accuses Anthropic of Inflating $30 Billion Run Rate by $8 Billion
OpenAI chief revenue officer Denise Dresser told staff Sunday that Anthropic's $30 billion run rate is overstated by roughly $8 billion. The fight is about gross versus net accounting on cloud-partner revenue, and both companies want to file S-1s this year.
Anthropic books the full amount billed through AWS, Azure, and Google Cloud as top-line revenue. OpenAI reports its Microsoft share net. Both methods comply with US GAAP. OpenAI's internal analysis puts Anthropic's comparable number closer to $22 billion, a delta large enough to reshape the IPO scoreboard months before a single prospectus is filed.
Timing tells the rest. Ramp's corporate-card data already puts Anthropic at 30.6% of AI-paying enterprise customers versus OpenAI's 35.2%, with the crossover projected within two months. Dresser's memo also concedes that the Microsoft partnership "has also limited our ability to meet enterprises where they are." Buried in paragraph three, that is not a victory lap.
Why This Matters:
- The gross-versus-net dispute lands on an auditor's desk the day Anthropic files its S-1. If the $30 billion headline gets restated, its $380 billion February valuation narrative follows.
- Enterprise buyers care less about accounting than product. Claude Code still draws the developer reverence that Codex cannot replicate, which is why Dresser is writing this memo to staff who read the same Ramp charts she does.
AI Image of the Day
Prompt: Two female gladiators facing each other, one human with worn leather armor, the other fully robotic with glowing neon pink elements, both holding a shared glowing heart between them, dark background, cinematic lighting, high contrast, ultra detailed anatomy and materials, emotional tension, modern dark art style, 8k resolution, masterpiece
Texas Man Charged With Attempted Murder in Molotov Attack on Altman Home
Federal and state prosecutors charged a 20-year-old Texas man Monday with attempted murder in Friday's Molotov cocktail attack on Sam Altman's San Francisco home. In his backpack: incendiary devices, a kerosene jug, and a written list of other AI executives by home address.
Daniel Moreno-Gama of Spring, Texas, allegedly threw a flaming bottle at the gate of Altman's Russian Hill residence around 3:45 a.m. Friday, then traveled to OpenAI's Mission Bay headquarters and tried to smash the glass with a chair. According to the federal affidavit, he told a security guard he had come to "burn it down and kill anyone inside." The state counts alone could carry 19 years to life.
The paper cops pulled off him is what turns this case into something bigger. Federal prosecutors describe a written statement claiming responsibility for attempting to kill Altman and urging others to do the same, with names and home addresses of other AI CEOs, board members, and investors attached. Court filings quote him writing about AI's "risk to humanity" and "our impending extinction." PauseAI, the Discord-based group opposing frontier AI development, said he joined its server two years ago and was banned Monday.
Why This Matters:
- The list changes the threat model for every frontier lab. Executive home addresses are already sitting in somebody's spreadsheet, and at least one person had clearly done the compiling before Friday.
- A second, apparently unrelated attack followed Sunday morning, when a gunman in a passing car fired at the same Russian Hill house. Two more suspects in custody. One incident is a warning. Two in three days is a pattern.
🧰 AI Toolbox
How to Hand Off Complex Research and Scheduling Tasks to an AI Super Agent with Genspark
Genspark is an AI workspace where a Super Agent plans multi-step tasks, picks the right tools, and delivers finished outputs instead of just answers. Ask it to compare competitors, build a slide deck, populate a spreadsheet with live web data, or make an actual phone call to book a restaurant. The system coordinates over 30 foundation models and 150 tools behind the scenes. Free tier includes 100 credits per day.
Tutorial:
- Go to genspark.ai and create a free account
- Open the Super Agent and give it a task you would normally spend 30 minutes on: "Research the top five project management tools, compare pricing and features, and put the results in a table"
- Watch the agent plan its approach, select tools, search the web, and compile structured results automatically
- Try AI Slides: type a topic and Genspark generates a full presentation deck with sourced content, then export to PowerPoint or PDF
- Use AI Sheets to auto-populate a spreadsheet with live data from the web, such as YouTube channel metrics or competitor pricing
- Test "Call for Me" by asking the agent to phone a restaurant and make a reservation on your behalf (available in the US, Canada, and Japan)
- Toggle "Mixture of Agents" in the chat to blend responses from Claude, ChatGPT, and Gemini into a single refined answer
What To Watch Next (24-72 hours)
- ASML Q1 earnings (Wednesday April 15, pre-market Europe): The Dutch lithography monopoly reports the first full quarter captured under the expanded US chip export regime. Bookings, not revenue, are the number to watch. A weak order book for EUV and High-NA tools signals that TSMC, Samsung, and Intel are pushing out capacity plans, and it reads straight across to 2027 AI accelerator supply. A strong book confirms the capex wave survives the tariff shock.
- TSMC Q1 earnings (Thursday April 16, Taipei): The most important AI chipmaker reports into a market that just watched Samsung project an eightfold profit jump. TSMC guided $34.6-35.8 billion in revenue, roughly 38% year-over-year growth. Management's commentary on Nvidia Blackwell yields, Apple silicon demand, and Arizona ramp costs will set the tone for every AI hardware name through earnings season.
- Semafor World Economy Summit (April 14-16, Washington): Global finance ministers, bank CEOs, and tech leaders converge alongside the IMF meetings for on-record sessions on AI capex, sovereign debt, and the tariff regime. Jamie Dimon, Jensen Huang, and European central bankers are on the program. This is the venue where private views on whether the AI build-out is a bubble get aired publicly, and any crack in the consensus tends to move markets the next morning.
🛠️ 5-Minute Skill: Turn a Wine List Into "Order This" in 30 Seconds
It is a client dinner in a city you do not know. The client hands you the wine list. Four different dishes are coming to the table. You have thirty seconds before the sommelier walks over, and you do not recognize half the producers.
Your raw input:
Spago, Beverly Hills. Table ordering Dover sole, Wagyu rib eye,
roast chicken, branzino. Five of us.
- 2019 Domaine Leflaive Puligny-Montrachet 1er Cru Les Pucelles, $340
- 2020 Château Musar Red, Bekaa Valley, $95
- 2018 Ridge Monte Bello, $420
- 2021 Gaja Ca'Marcanda Promis, Tuscany, $140
- 2019 Domaine Tempier Bandol Rouge, $120
The prompt:
Pick one white and one red that (1) pair with more than one dish
on the table, (2) sit in the middle of the list on price, (3) are
producers a serious wine drinker would respect. One line for why
each. One line I can say to the sommelier without sounding rehearsed.
The output:
- White: Leflaive Puligny-Montrachet Les Pucelles. Covers the Dover sole and the branzino, top Burgundy house, nobody at a serious table will second-guess it. Say: "The Leflaive, it works with both fish."
- Red: Domaine Tempier Bandol Rouge. Handles the rib eye and the roast chicken, and Bandol is the pick of people who drink wine for a living, not the pick of people trying to look like they do. Say: "Let's do the Tempier."
Why this works:
You are not asking the AI to teach you wine. You are asking it to narrow a list you already have in front of you, against a specific table order and a social constraint. That is the part that is actually hard in thirty seconds, and it is the part Google cannot answer.
What to use:
Snap the wine list with Claude or ChatGPT mobile, paste the orders underneath, run the prompt. Claude is slightly more honest about producer reputation. ChatGPT is faster on mobile. Save the prompt and reuse it at every client dinner for the rest of the year.
📖 AI Alphabet
|
E
|
📖 AI Alphabet Embedding An embedding turns text, images, or other data into numbers that capture meaning. That makes it easier for software to search, compare, and group similar information. |
AI & Tech News
Iran Hits 45-Day Internet Blackout, Cutting 90 Million From Global Web
Iran's near-total disconnection from the global internet reached 45 consecutive days, the longest sustained blackout NetBlocks has tracked, with citizens routed entirely through the state-controlled National Information Network. The isolation follows escalating regional strikes and hands Tehran full control over what its population sees, searches, and transmits.
Anthropic Keeps European Regulators Out of Mythos Model Testing
Anthropic restricted early access to its new "Mythos" model to a small circle of U.S. companies and the UK's AI Safety Institute, leaving EU regulators without formal notification or evaluation before the limited rollout. Brussels officials say the exclusion undermines coordinated oversight at a moment when Anthropic itself warns Mythos raises meaningful cybersecurity risks.
CodeWall Breaches Bain's Internal AI Tool Using Exposed Credentials
AI penetration-testing firm CodeWall said it broke into Bain & Company's internal AI system, Pyxis, by exploiting credentials leaked in public web code, the second such breach at a top consulting firm after a March intrusion at McKinsey. The back-to-back incidents expose how quickly enterprise AI tools accumulate attack surface as firms race to deploy them.
Amazon Nears Globalstar Acquisition to Challenge Starlink
Amazon is in advanced talks to buy satellite operator Globalstar, with a deal potentially announced as soon as Tuesday, aiming to accelerate its Project Kuiper broadband ambitions against SpaceX's Starlink. Globalstar shares jumped more than 15% in pre-market trading on the Bloomberg report.
Oracle Commits to 2.8 Gigawatts of Bloom Energy Fuel Cells
Oracle expanded its power deal with Bloom Energy to procure up to 2.8 gigawatts of on-site fuel cell capacity, days after receiving a warrant to buy nearly $400 million in Bloom stock. The arrangement locks in behind-the-meter generation for Oracle's data center buildout and gives the cloud provider financial upside if Bloom's shares continue climbing.
China's YMTC Plans Two New Fabs to Double NAND Capacity
Yangtze Memory Technologies intends to build two additional memory fabs beyond its first plant scheduled for 2026 completion, more than doubling the Chinese chipmaker's output capacity. The expansion pushes Beijing's bid for NAND self-sufficiency forward despite U.S. export controls meant to slow exactly this kind of scaling.
Handshake and Mercor Each Cross $1 Billion in Annualized Revenue
Data labeling startups Handshake and Mercor each passed $1 billion in gross annualized revenue as AI companies burn through demand for human-labeled training data, with Handshake jumping from $550 million in January. The numbers reframe data contractors as tier-one AI infrastructure rather than back-office support.
Missouri Town Ousts Entire City Council After Data Center Vote
Voters in Festus, Missouri, unseated all four incumbent council members up for reelection days after the council approved a $6 billion data center on a high-turnout protest vote. The recall-style sweep is a warning to every local government weighing tax incentives against community backlash as hyperscaler projects hit small towns.
🚀 AI Profiles: The Companies Defining Tomorrow
Handshake spent 12 years as a job board for college graduates. Then it pivoted into AI data labeling and hit a $1 billion revenue run rate in a year. 🎓
Founders
Garrett Lord, Ben Christensen, and Scott Ringwelski founded Handshake in 2013 at Michigan Technological University, where they built a recruiting platform aimed at students outside the Ivy League circuit. Headquartered in San Francisco, the company employs roughly 550 people after cutting 100 positions (15% of its US recruiting workforce) earlier this year in what Lord called a "re-founding" around AI.
Product
Handshake AI, launched in January 2025, sells expert human feedback to frontier AI labs training reasoning and reinforcement learning systems. The wedge is the existing user base: 500,000 PhDs and 3 million master's students already on the platform, pre-vetted by field, willing to label complex math, biology, and law tasks that generalist annotators cannot handle. Revenue moved from roughly $5 million a year ago to $550 million in January to a $1 billion annualized run rate this month, per board member Mamoon Hamid. The AI arm grew from 15 people to 150 in its first eight months and now exceeds the legacy recruiting business.
Competition
Scale AI (partially absorbed by Meta), Surge AI, Mercor, and Turing all chase the same frontier-lab contracts. Handshake's moat is the PhD roster it already owns, which rivals have to acquire customer by customer.
Financing 💰
Total funding around $435 million from Coatue, Kleiner Perkins, Spark Capital, EQT Ventures, GGV Capital, Chan Zuckerberg Initiative, and Base10 Partners. Last disclosed valuation sits at $3.5 billion, a mark that now looks stale given the AI revenue trajectory.
Future ⭐⭐⭐⭐
The pivot is the rarest kind in software, a legacy marketplace that turned its user graph into a defensible AI training asset in under a year. The risk is concentration: a handful of labs drive most of the revenue, and any one of them building an in-house annotation stack would dent the growth story. Watch for a new funding round priced off the $1 billion run rate, not the 2025 valuation. 🎓
Our Rating: ⭐⭐⭐⭐ (out of 5)
🔥 Yeah, But...
OpenAI CRO Denise Dresser sent a four-page memo to staff on Sunday claiming Anthropic inflates its $30 billion run rate by $8 billion by counting gross AWS, Azure, and Google Cloud billings rather than net revenue. The same memo concedes OpenAI's Microsoft partnership has "limited our ability to meet enterprises where they are." Ramp data shows Anthropic on track to overtake OpenAI in enterprise paying customers within two months.
Sources: Implicator.ai, April 13, 2026
Our take: At an $852 billion valuation, you would think OpenAI could afford to be relaxed about how the runner-up fills out its revenue line. Instead, the sales chief spent her Sunday writing four pages on GAAP treatment, which is roughly the energy of a Michelin chef storming out of his kitchen to critique how the bistro next door plates its fries.
Both methods comply with the rules. The quieter admission, buried in paragraph three, is that the Microsoft partnership has stopped OpenAI from selling to the customers it wants. She is right. She is also describing it in a document about someone else's math.
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: [email protected]
