The Operator’s Lens

Welcome to August’s growth newsletter, where we track what matters in late-stage venture.

Last month we said a ceiling was forming at the top. August kept the M&A wave rolling. SpaceX closed its $60B Cursor deal. Stripe dropped $7.5B on OpenRouter. Walmart grabbed Vibe.co for $1.4B. Q2 was already the strongest quarter for venture-backed exits on record - August showed no signs of slowing down.

What’s worth noting: none of these buyers acquired a model. They acquired distribution - developers, API traffic, ad spend. Whether that’s a pattern or a coincidence, it’s where the money moved this month.

The IPO window stayed open for biotech - three VC-backed names raised $870M+ in the first week alone - but quiet for tech. Anthropic’s ARR jumped from $47B to $65B in a single month and is targeting an October debut. We’ll see if public markets agree with the private marks.

Let’s get into it.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $1B for Valar Atomics, an El Segundo, CA-based compact nuclear reactor company for AI data centers, with funding from Sequoia Capital, Apandion, and Point72.

• $400M for Higgsfield, a San Francisco, CA-based AI-native video and image creation platform, with funding from DST Global, Tribe Capital, and Accel.

• $280M for Wispr Flow, a San Francisco, CA-based AI speech-recognition and dictation software company, with funding from Menlo Ventures, Notable Capital, and NEA.

• $136M for Point2 Technology, a San Jose, CA-based RF interconnect solutions company for AI data centers, with funding from LB Investment, Arm, and NVIDIA.

• $57M for Convex, a San Francisco, CA-based reactive backend platform for software development, with funding from Insight Partners, Andreessen Horowitz, and Spark Capital.

Series C

• $1B for Castelion, a Torrance, CA-based hypersonic strike weapons manufacturer, with funding from JPMorgan Chase, Andreessen Horowitz, and Carlyle.

• $250M for Neros Technologies, a Torrance, CA-based autonomous strike drone and interceptor manufacturer, with funding from Sequoia Capital, American Strategic Technology Fund, and Spark Capital.

• $250M for Muon Space, a Mountain View, CA-based satellite constellation company, with funding from Eclipse Capital, Google, and Salesforce Ventures.

• $150M for HappyRobot, a San Francisco, CA-based agentic AI platform for enterprise logistics operations, with funding from Prysm Capital, Eurazeo, and Andreessen Horowitz.

• $100M for Rillet, a San Francisco, CA-based AI-native accounting and ERP platform, with funding from ICONIQ, Sequoia Capital, and Andreessen Horowitz.

Series D

• $1.37B for Hadrian, a Torrance, CA-based AI-powered automated factory company for aerospace and defense, with funding from WCM Investment Management, JPMorgan Chase, and Andreessen Horowitz.

• $700M for Etched, a San Jose, CA-based AI inference chip startup building transformer hardware, with funding from Jane Street, Kleiner Perkins, and Sequoia Capital.

• $85M for Obsidian Security, a Palo Alto, CA-based runtime security platform for enterprise applications and AI agents, with funding from Crescent Cove Advisors, Greylock Partners, and Menlo Ventures.

• $80M for MarginEdge, an Arlington, VA-based restaurant back-office management SaaS platform, with funding from Schooner Capital, Ten Coves Capital, and Osage Venture Partners.

• $66M for Abcuro, a Newton, MA-based clinical-stage biotech developing immunotherapies for rare autoimmune diseases, with funding from New Leaf Venture Partners, abrdn, and Bain Capital Life Sciences.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

The top two are still frozen at $1.8T combined — but they’re finally growing into it. Anthropic’s ARR jumped from $47B to $65B in a single month. OpenAI nearly doubled to $40B+. For the first time, the mega-marks are starting to look less insane. The real story is at the bottom: Shein priced its HK IPO at $27B — a 60% haircut from its $66B private mark. That’s the cost of going public when the narrative runs out. Anduril slides in at #10, swapping fast fashion for fighter drones.

How to Play It

ByteDance at 3.2x revenue is still the cheapest name on a list of 20x+ multiples — and it just leapfrogged Tether for #3. Databricks posted 80% ARR growth at $7B+ scale; at $190B, it's the one name where the valuation and the fundamentals are moving in the same direction. The Shein markdown is a warning shot for anyone holding late-stage paper at 2021 prices. Public markets aren't giving it back.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• Anysphere (Cursor) - acquired by SpaceX for $60 billion

• OpenRouter - acquired by Stripe for approximately $7.5 billion

• Vibe.co - acquired by Walmart for $1.4 billion

• NextSlide - acquired by OpenAI for an undisclosed amount

• Lica - acquired by Gamma for an undisclosed amount

Top Venture-Backed IPOs

• Braveheart Bio (BRVE) | San Francisco, CA-based clinical-stage cardiovascular biotech developing treatments for hypertrophic cardiomyopathy, backed by Andreessen Horowitz, OrbiMed, and Forbion. Raised $382.5M in IPO

• Latigo Biotherapeutics (LTGO) | Thousand Oaks, CA-based clinical-stage biopharma developing non-opioid pain treatments, backed by Westlake Village BioPartners, 5AM Ventures, and Foresite Capital. Raised $346M in IPO

• BlossomHill Therapeutics (BLSM) | San Diego, CA-based precision oncology company developing small-molecule cancer therapies, backed by Colt Ventures, Vivo Capital, and OrbiMed. Raised $150M in IPO

This Month at R136 Ventures

That’s all for this month. See you in September!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

In 2012, I watched a bank nearly die.

I was CTO for a financial institution that processed half of a country’s transactions. Over 20,000 engineers reported up through my org, and the IT budget ran to roughly $5 billion a year.

One morning, during a large database migration, the core system crashed. Cards stopped working at gas stations, ATMs went dark, and millions of people were locked out of their money. It made national television within an hour.

My senior engineering team asked for four hours to run safe diagnostics and restore from backup (which was standard procedure). But I knew that a four-hour blackout for a bank that half a country depended on would permanently destroy its reputation. Customers don’t come back after that.

Then an engineer named Alex proposed something different. A risky, unverified patch that had never been tested in production. If it worked, recovery in minutes but if it failed, it could wipe the ledgers entirely…maybe crash the system for a week.

I told Alex to push the button.

Luckily it worked and the system came back in minutes.

Afterward, I called my CEO fully prepared to resign and take accountability. His response: “Shut up... just go and solve the problem.”

I tell this story because it shows what people outside banking consistently underestimate. Banks are survival machines. Every compliance department, risk committee, licensing regime exists for one reason: these institutions hold society’s money, and when they fail, the damage spreads far beyond their walls.

That survival instinct is baked into everything they do. Every process, every approval layer, and audit trail assumes a human is on the other end, accountable for what happens next.

Now think about what the agent economy actually requires. We are asking these same institutions to let autonomous software agents move real money through their systems. Billions of agents, transacting by the millisecond, forming temporary relationships with other agents and dissolving them just as fast. No human in the loop for most of it.

Here’s where two problems collide. The payment infrastructure those agents would run on, Visa, Mastercard, ACH, SWIFT, even Stripe, was built assuming a human initiated every transaction. It physically cannot handle machine-speed volumes.

And at the same time, the institutions that control those rails have an entire risk culture built around human accountability. Someone signs, approves and is responsible when things go wrong. But agents don’t fit that model.

So you can’t solve one without the other. Faster pipes alone don’t work if the banks operating them have no way to trust what’s flowing through. And institutional trust alone doesn’t work if the rails underneath collapse under the volume.

Both problems have to be solved together before the agent economy can function at any real scale.

The pipes were built for humans

Everyone in venture capital is chasing LLM benchmarks and GPU clusters right now. Fair enough. But underneath the intelligence race, there’s a structural problem forming that most people aren’t paying attention to.

The scale is hard to estimate. McKinsey projects that AI agents could orchestrate $3 trillion to $5 trillion in global B2C retail revenue by 2030. Juniper Research forecasts $1.5 trillion in agentic commerce spend by the same year. The gap between those numbers is itself the most useful data point. Nobody actually knows but even the low end demands infrastructure that just doesn’t exist today.

Here’s the problem: an AI agent can compare 10,000 offers, negotiate terms, and select the optimal vendor in seconds.

Then it hits a payment wall so it needs a human’s card, a human’s approval and a human’s identity. The agent that just outperformed an entire procurement team sits idle, waiting for someone to type in a credit card number.

Most people think a payment is one event but it’s actually three separate processes: authorization, clearing, and settlement. They happen at different times, through different systems, governed by different rules. When you tap your card at a coffee shop, the authorization feels instant. The actual movement of funds between banks settles hours or days later in batches.

That batching made sense for decades. Human transaction volumes are pretty manageable. They might make ten purchases a day, each worth several dollars, each producing behavioral signals that fraud systems can evaluate. Location, typing speed, purchase history. The interchange fee on each transaction is easily absorbed because the transaction itself is large enough.

But now imagine an enterprise AI agent tasking 50 specialized sub-agents with writing, compiling, testing, and deploying a software patch in real time. To complete this, the primary agent needs to hire, negotiate, and instantly pay those fifty agents for compute loads measured in milliseconds.

Under current rails, a $0.0001 micropayment gets swallowed by a transaction fee that can be multiples of the payment itself. You’re paying more to move the money than the money is actually worth.

ACH processed $104 trillion in 2024, representing 74% of all U.S. noncash-payment value. That’s an extraordinary system. But Same Day ACH operates on three settlement windows during business hours and shuts down on weekends and federal holidays. If your agent fleet needs compute resources at midnight on a Sunday, too bad.

Swift moves 13.4 billion messages a year. But Swift is a messaging layer. The actual friction comes from the correspondent banking chain underneath: multiple intermediary banks, each adding compliance checks, each operating in different time zones, each holding funds in transit. For a cross-border agent transaction that needs to settle in milliseconds, that architecture adds days.

These systems are genuine engineering achievements, they work at enormous scale. But they were optimized for a world of larger, less frequent, human-authorized transactions. An agent making 10,000 micro-purchases per hour just wasn’t in the design spec.

What invisible infrastructure actually looks like

In 2001, long before the iPhone existed, my team built a mobile banking system in Eastern Europe. Phones were expensive and clunky and with primitive internet connectivity. So we built a system that ran entirely on SMS text messages.

A user would type something like “Send 500 rubles to +7-123-456-7890 for dinner” and money moved. No ap or login screen or dashboard.

What surprised us is that users hijacked the system to flirt, to communicate, to have conversations where the payment was the excuse and the message was the point. Money became a language.

I think back on this because it taught me something I think most people in tech still underestimate. We succeeded because we relied on communication, on the simplest possible interface.

We weren’t actually building a banking app. We were building a way for people to talk to each other using money. And we hid the massive database complexity of early 2000s banking systems behind a twelve-word text message. A million people used it without thinking about what was happening under the surface.

The lesson for the agentic era is the same. The best infrastructure is invisible.

Tomorrow’s agentic rails need to hide the complexity of machine-to-machine billing the same way that SMS system hid the banking stack. The human user shouldn’t see individual ledger settlements, clearing pipelines, or API calls. The agents should communicate, transact, and settle autonomously in the background. Value flowing as seamlessly as electricity.

The protocol race is already underway

The biggest companies in payments are now building this. They’re further along than most people realize.

In the last 18 months, every major payments player launched an agentic commerce initiative. Nobody is building the same thing. Think of it as layers of a new stack. At the bottom, MCP connects agents to APIs and data sources. One layer up, Google’s A2A protocol lets agents discover and coordinate with each other across platforms.

OpenAI and Stripe built ACP to structure how agents browse catalogs, manage carts, and check out. Google separately launched UCP to cover the full commerce lifecycle from product discovery through post-purchase support.

Then there’s the authority layer (this is where things get genuinely interesting). Google’s AP2 protocol, backed by more than 60 organizations including PayPal, uses cryptographically signed mandates to create provable evidence that a human authorized a specific transaction under specific constraints.

This is the piece that matters most for the trust problem I described earlier. It’s scoped, time-limited, auditable delegation. Not “give the machine your credit card.”

At the execution layer, Stripe’s Machine Payments Protocol uses HTTP 402 payment challenges for machine-to-machine purchases. When one agent needs to buy a resource from another, the server responds with a payment challenge instead of an error page. The buying agent resolves it programmatically. No checkout flow. No cart. Just a protocol-level exchange that settles the transaction.

The incumbents are moving fast. Mastercard launched Agent Pay in April 2025, then expanded with Agent Pay for Machines in June 2026, supporting micro-value, high-frequency transactions across cards, bank accounts, and stablecoins. Visa is developing agent-identification mechanisms.

PayPal released an Agent Toolkit and joined AP2. Santander and Mastercard completed Europe’s first live end-to-end payment executed by an AI agent. Robinhood opened its platform to AI agents in May 2026.

Nobody owns the full stack. These protocols overlap, compete, and will probably coexist for years. The likely outcome is gateways that translate between them. And that translation layer, the middleware for non-human commerce, is itself an enormous opportunity.

Trust is the real infrastructure

Protocols solve the conversation between agents and merchants. But the hardest question isn’t technical.

When an agent buys the wrong thing, who absorbs the loss? Maybe the user gave an ambiguous mandate, or the agent misinterpreted it. Maybe a merchant manipulated the agent’s decision-making or credentials were stolen.

Card chargebacks protect consumers today, but they were designed for a human calling their bank to dispute a charge. Not for machines generating thousands of micro-exceptions per hour where the investigation cost for each one exceeds the transaction value.

Cryptographic mandates like AP2 improve the quality of evidence. They create a provable record of what the human authorized. But they don’t by themselves settle legal liability when the agent’s interpretation diverges from what the human actually intended.

I spent 20 years inside banking systems. The real infrastructure underneath any payment system is the liability framework which answers the question - who bears the loss when something goes wrong. Every payment system in history had to answer this question before it could scale. Agent payments won’t be different.

The trillion-agent economy cannot exist until we solve a specific problem: strict, real-time auditability built directly into the ledger. Every transaction traceable, every agent accountable, every mandate provable.

Because when the system crashes (and at some point it will) someone has to explain exactly what happened and why. That someone can’t be an AI that hallucinates its own audit trail.

Where the value concentrates

The AI models themselves are becoming cheaper fast. The value is moving to the layers above them. I see it concentrating in three places.

Identity and authority

The core question: how do you verify who an agent is and what it’s allowed to do? And can you prove that after the fact? The U.S. National Institute of Standards and Technology is already working on standards for how agents identify themselves and get permission to act.

Google’s payment authorization protocol and Mastercard’s system for proving what a human actually approved are early answers. Most large companies already have far more machine accounts than human ones. Tens of thousands, sometimes millions. And the identity systems we rely on were built for people with usernames and passwords. That gap has to close.

Settlement modernization

Money needs to move instantly, around the clock, with no batch processing delays. Several approaches work for different situations. Stablecoins handle cross-border payments and recurring digital services well. The GENIUS Act, signed July 2025, creates the first real U.S. federal framework for stablecoins used as payment.

The full operational rules don’t kick in until January 2027, but the legal direction is already set. Instant bank transfers cover other flows. And for big institutional deals where multiple parties need to settle privately and in sync, specialized networks are being built. The right answer is probably all of them, coexisting.

Middleware and orchestration

I already described how many competing protocols already exist. So somebody has to build the connectors between them. The gateways that let an agent using one protocol talk to a merchant using another, and settle the payment across whatever rail makes sense. Think of it as the translation layer for non-human commerce. Given how fragmented things already are, whoever builds that sits in a very valuable position.

What I see up close: Airwallex is building AgentOS, a set of tools that lets AI agents do real financial operations. Check balances. Send payments. Create invoices. Issue virtual cards. Handle currency conversions across accounts. The interesting part is the guardrails they built in.

Anything that actually moves money, transfers, payouts, currency exchanges, requires human approval by default. An agent can look up information and prepare transactions all day. But before money leaves an account, a person has to sign off.

That’s what responsible agent finance looks like at this stage. Let the machines handle the complexity, keep humans in control of what matters. (Disclosure: R136 Ventures is an investor in Airwallex.)

Digital Asset’s Canton Network solves a different piece. Canton is a settlement network built for institutions. Its key feature: when two sides of a deal are linked, both go through or neither does. No half-completed transactions.

If I’m selling you a bond and you’re paying me cash, both legs happen together, or the whole thing rolls back. That sounds simple but matters enormously when agents start creating thousands of linked deals per second across systems that don’t naturally talk to each other.

Broadridge runs hundreds of billions in daily repo transactions on Canton. DTCC partnered with Digital Asset to turn U.S. Treasuries into digital tokens on the network. When the volume of linked obligations outpaces what today’s separate record-keeping systems can handle, synchronized all-or-nothing settlement becomes the only way forward. (Disclosure: R136 Ventures is an investor in Digital Asset.)

When owning and spending become the same thing

Follow this thread to its end and something bigger takes shape.

Real-world assets are gradually being turned into digital tokens. Stocks, real estate, intellectual property. When that process reaches sufficient scale, the line between owning something and spending it starts to disappear.

An AI agent won’t need to sell a stock position, wait for settlement, transfer the cash to a bank, and then send a payment. The agent will instantly swap a fraction of a tokenized S&P 500 share to pay another agent for a millisecond of computing time.

That’s what full financial velocity looks like. Everything becomes liquid. Every asset can be exchanged directly for any other asset, settled in the same moment. The old distinction between money and property fades into a single layer where everything is programmable and everything moves.

We’re not there yet. And maybe we wont be for years but the regulatory framework is forming. The GENIUS Act and the EU AI Act’s transparency obligations. The institutional plumbing is being laid and the protocols keep multiplying.

But the direction is clear.

The companies building the identity layer, the settlement layer, and the liability framework for machine transactions right now are building the rails that everything else will run on.

I learned this in 2001, watching a million people send money through twelve-word text messages on Nokia phones. I learned it again at 3 AM in 2012, telling Alex to push the button on a patch that had no right to work. The technology changes every few years. The plumbing always matters more than the interface. That hasn’t changed once.

The Operator’s Lens

Welcome to July’s growth newsletter, where we track what matters in late-stage venture.

Last month we said the divide is the market structure., and July made it even weirder. Anthropic and OpenAI both held flat..nearly $3 trillion combined just sitting there while Databricks surged 40% to $188B.

But the real story this month landed just last week when Moonshot dropped Kimi K3, an open-weight model that matches frontier closed systems on coding benchmarks. Within a week 50+ companies signed a letter to DC arguing open weights are critical for American AI leadership. OpenAI signed it, a company that built its entire brand on closed architecture, now pushing for the thing that could eat its own pricing power.

Anthropic didn't sign. The most valuable private company on earth at $965 billion sat it out while its biggest competitors rallied around open weights. If the models all converge, value moves to infrastructure and distribution.

Let’s get into it.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $180M for Augustus, a Dallas, TX-based fintech with funding from Tiger Global. Hummingbird and QED Investors.

• $91M for Venus Aerospace, a Houston, TX-based hypersonic propulsion technology company, with funding from Mercury Fund, Lockheed Martin Ventures, and MESH.

• $85M for Rylo, a New York, NY-based AI communication accessibility platform, with funding from Canaan, General Catalyst, and Vertex Ventures.

• $70M for Flex, a Miami, FL-based AI banking platform, with funding from Crosslink Capital, FJ Labs, and Portage.

• $55M for Bunkerhill Health, Palo Alto-based AI health-tech company with funding from Sequoia Capital, Felicis, Optum Ventures and others.

Series C

• $800M for Together AI, a San Francisco, CA-based open-source AI infrastructure platform, with funding from Aramco Ventures, Vista Equity, General Catalyst, Nvidia, and Salesforce Ventures

• $400M for Chai Discovery, a San Francisco, CA-based AI-powered drug discovery company, with funding from Index Ventures, Kleiner Perkins, and Sequoia Capital.

• $300M for Etched, a San Jose, CA-based AI chip company with funding from Sequoia Capital, Andreessen Horowitz and SK Hynix.

• $130M for Emergent Labs, an Oakland, CA-based AI developer platform, with funding from Creaegis, Claypond Capital, and Khosla Ventures.

• $120M for Norm Ai, a New York, NY-based legal AI platform, with funding from Khosla Ventures, Blackstone, and Bain Capital Ventures.

Series D

• $1.505B for Fireworks AI, a Redwood City, CA-based enterprise AI inference and customization platform, with funding from Atreides Management, Index Ventures, and TCV.

• $650M for Wonder, a New York, NY-based food delivery and kitchen operator, with funding from Accel, GV, and New Enterprise Associates (NEA).

• $315M for AdvanCell, a Boston, MA-based clinical-stage radiopharmaceutical company, with funding from Ally Bridge Group, Alpha Wave Global, and Bain Capital Life Sciences.

• $120M for Candid Health, San Francisco, CA-based AI-native revenue cycle management company with funding from Sith Street Growth, 8VC, Y Combinator, and others.

• $100M for Spectro Cloud, a San Jose, CA-based Kubernetes infrastructure management platform, with funding from Goldman Sachs Alternatives, AMD Ventures, and Ericsson.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

The top two haven't moved in two months - $1.8T combined, frozen. Meanwhile Tether added $150B on vibes alone, ByteDance shed 20% on a secondary reprice, and Databricks jumped 40% on an actual round. This list is splitting into two markets: AI names trading on narrative, everything else trading on cash flow.

How to Play It

ByteDance at 2.6x revenue and Shein at 1.5x are the two cheapest names on a list dominated by 20x+ multiples - if you're buying secondaries, that's where the margin of safety lives. Databricks just proved a single strategic round can move a $130B company 40% overnight. The mid-list is where the upside trades now, not at the frozen top.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• Fin - acquired by Salesforce for $3.6 billion

• Bud (formerly Orchids) - acquired by Figma for an undisclosed amount

• zio AI Corporation (Azio AI) - acquired by Envirotech Vehicles, Inc. for $750 million

• ​Shaped - acquired by Whatnot for an undisclosed amount

• ​SigScalr, Inc. - acquired by Apple Inc. for an undisclosed amount

Top Venture-Backed IPOs

• Lime (LIME) | San Francisco, CA-based shared e-bike and scooter micromobility company backed by Bain Capital Ventures, Andreessen Horowitz, and GV. Raised $167M in IPO

• Scribe Therapeutics (SCTX) | Alameda, CA-based clinical-stage biotech developing CRISPR-based genetic medicines backed by Andreessen Horowitz, OrbiMed, and Perceptive Advisors. Raised $128.7M in IPO

This Month at R136 Ventures

That’s all for this month. See you in August!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

Every year the same vendors came in with the same pitch: you’re growing, you need more seats. And it was true. We hired people, those people needed tools, we bought licenses. The vendor’s revenue scaled with our headcount. Expansion revenue accounted for 40% to 67% of new ARR for SaaS companies at scale. Nobody questioned it because it worked.

I think about that math a lot now. Because from where I sit today, as an investor backing software companies, I can see both sides of a problem that is just starting to become visible. The unit that the entire software industry was built on, the human seat, is losing its meaning.

The most productive user of that software no longer needs a login and every vendor who built their financial model on counting humans in front of screens is about to learn what happens when those humans step away but the work keeps getting done.

The billing unit broke

Here is what is actually happening inside enterprises right now.

Companies are deploying AI agents across support, sales operations, and back-office workflows. And they are seeing that entire functions can run with 20% to 90% fewer human seats after those agents go live. The work still gets done (and often better) but the humans aren’t logging in.

I know exactly what happens next because I used to be the buyer. You open your renewal spreadsheet, you see 3,000 licenses, then you check the usage data. 1,200 people haven’t logged in for ninety days. Your procurement team calls the vendor. The conversation is uncomfortable for everyone.

Multiply that across thousands of enterprises and you get what the market delivered in early 2026: an estimated $1 to $2 trillion in software market capitalization, gone. Analysts called it the “SaaSpocalypse.” The name sounds dramatic but the cause was simple. Investors realized that per-seat revenue was structurally compromised.

No one has figured out how to price AI software yet

When I was the buyer, I never questioned per-seat pricing bc it felt fair. You use the tool, you pay for the tool…simple. But fairness depended on the assumption that a human being would always be the one using it.

Remove that assumption and the whole contract structure starts to feel arbitrary. You are paying for access to system that your AI agent consumes through an API. The agent doesn’t need the interface. It doesn’t need onboarding. It doesn’t need a password reset at 2 AM. Why does it need a seat?

This is the question every enterprise procurement team is now asking and no vendor has a good answer yet. Some are trying.

• Intercom charges $0.99 per support ticket its AI agent resolves

• Zendesk charges $1.00 to $1.50 per automated resolution.

• Salesforce now runs three different pricing models simultaneously for Agentforce. Three models at once. And it’s probably the right posture when you don’t know what comes next.

Outcome-based pricing sounds clean in press release but in practice it’s really hard. You need to define what “resolved” means in a contract that will survive your customer’s legal team at renewal.

I have been on that side of the table and know what happens when a vendor promises an outcome that procurement can interpret six different ways. It creates a new category of negotiation that most sales organizations are not ready for.

Consumption-based pricing, pay for API calls or tokens processed, has opposite problem. It is transparent. Customers like it until they get a bill from AI processes nobody was monitoring. And vendors lose the revenue predictability they spent a decade building.

The finance team that used to forecast next quarter with 95% accuracy now stares at a usage curve that could go anywhere.

Hybrid models, a base platform fee plus variable AI charges, are where most large vendors have landed. ServiceNow replaced its legacy pricing with AI-native tiers. Workday launched Flex Credits. Everyone is hedging. Nobody has arrived.

My honest read - the industry is in the early phase of a transition that will take years to settle. The companies pretending they have the answer are probably the ones most at risk of getting it wrong. Running three pricing models at once tells you where the industry stands. Everyone is searching. Nobody has found it.

The infrastructure layer is where the real value will be built

And the pricing question sits on top of a deeper structural problem that most people are not talking about yet. AI agents don’t click buttons, they call APIs. Software that was designed for human eyes and human hands is being bypassed entirely.

89% of developers now use AI in their work, but only 24% are designing APIs specifically for AI agent consumption. That gap is enormous and tells you where the next generation of infrastructure companies will be built.

Anyone who has managed large-scale enterprise IT knows what happens when systems can’t talk to each other. I lived that reality for years across dozens of platforms. The cost is staggering and mostly invisible until you try to do something that crosses system boundaries.

Now imagine that problem multiplied by thousands of autonomous agents, each needing credentials, permissions, guardrails, and the ability to communicate with other agents across your entire infrastructure.

This is the part that keeps me up at night as an investor. The identity problem. Machine identities already outnumber human users by orders of magnitude in most enterprises. Every agent needs access policies.

Traditional identity management was built for people with logins and passwords. It is fundamentally inadequate for governing fleets of autonomous software agents making API calls across your entire stack.

When I was running technology for a bank that processed a significant share of an entire country’s transactions, interoperability between systems was already one of the hardest problems we faced. And that was with humans driving the workflow.

Add thousands of autonomous agents operating at machine speed, each with its own credentials and permissions, and you start to understand why the infrastructure layer underneath all of this is where the real value will be created. The plumbing. The boring, essential, deeply defensible work that nobody talks about at conferences.

Fifty-seven percent of companies had AI agents running in production by late 2025. For large enterprises with over 10,000 employees, that number was 67%. These are not pilots. These are production systems handling real workflows. And most of them are operating on infrastructure that was never designed for this kind of load.

Where humans still matter and why there will be fewer of them

Before you assume the human seat is dead everywhere, it’s worth pausing. In regulated industries like finance, healthcare, legal, accountability still requires a person. An AI agent can prepare 99% of a compliance review, but someone with a license and legal liability has to sign off. I dont think that’s going away.

And strategic, creative, relational work resists automation in ways that are genuinely structural. The C-suite executive, the design lead, the key account manager. Their value comes from judgment and relationships, not from workflows that can be scripted.

But the nature of oversight is changing. The old model, where a person manually approves every AI action, creates bottlenecks and leads to rubber-stamping. People get desensitized and stop genuinely reviewing.

The smarter approach is tiered: low-risk tasks run fully automated, high-stakes decisions require mandatory human approval. Fewer humans, more powerful roles. More expensive seats, but far fewer of them.

I managed teams large enough to see this pattern. The most valuable person is rarely the one doing the most work. It is the one making the hardest calls with the least information. AI doesn’t change that. It amplifies it.

The per-seat era is over and three models will replace it

So what does all of this add up to?

The entire SaaS era was built on elegant assumption that the number of people using software was a good proxy for the value it delivered. And for thirty years that assumption held while creating enormous wealth for the companies and investors who understood it.

It doesn’t hold anymore.

At its core, this is business model story. The technology is the catalyst, but the real disruption is in how value gets measured, priced, and captured. Every pricing experiment, every valuation reset in the public markets, every startup building agent governance infrastructure. It all traces back to one structural fact: the billing unit changed.

1. Some companies will figure out how to charge for outcomes.

2. Some will build the infrastructure that agents depend on.

3. Some will hold onto seats in domains where human judgment is genuinely irreplaceable.

I think those are the three paths forward.

I spent years buying software for one of the largest tech organizations in European banking. I know what makes a vendor irreplaceable at renewal and what makes them a line item that gets cut without a second thought.

The irreplaceable ones were never the prettiest tools but the ones so deeply embedded in our workflows that ripping them out would cost more than keeping them in. They were the plumbing. The rails, not the trains.

That has not changed. What has changed is that the workflows themselves now run on machine logic instead of human habits. The vendors who understand this will reprice, rebuild, and survive. The ones who keep counting seats that nobody is sitting in will slowly compress until there is nothing left to compress.

The billing unit changed and the strange thing is, the software itself is fine. It still works. It still solves problems. It might even solve them better now than it did a year ago. The only thing that broke is the assumption about who would be using it.

Thirty years is a long time for an assumption to hold. Long enough that people stopped seeing it as an assumption and started treating it as a fact.

That is always how these transitions catch people off guard. The fact was never a fact. It was a convenience but now the convenience is over.

In this conversation, we discuss:

• Why sending money home can be a life-or-death issue

• What founders should understand about Africa and Nigeria

• Why Nigeria is one of the most important fintech markets

• How stablecoins reduce the cost of cross-border payments

• Why Zuba is focused on affordability, speed, and reliability

• How AI and blockchain can help fight financial fraud

• Why customer care still matters in the age of AI

• What it takes to build a global fintech company

• How Josiah thinks about talent, focus, discipline, and obsession

This episode is brought to you by R136 Ventures, supporting visionary founders building the next generation of fintech, B2B, and AI-driven companies.

The Operator’s Lens

Welcome to June’s growth newsletter, where we track what matters in late-stage venture.

Last month we said only hard problems get liquidity. Turns out the market heard that and decided to price accordingly. SpaceX went public at $1.75 trillion (biggest IPO ever) and Anthropic leapfrogged OpenAI to become the most valuable private company on earth at $965 billion. The divide isn't narrowing but becoming the entire market structure. So if you're not reshaping a cost curve or owning an infrastructure layer, you're not just getting marked down..you're getting ignored.

Let’s get into it.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $12B for Prometheus, a San Francisco, CA-based AI for engineering and manufacturing company, with funding from JPMorgan, BlackRock, and Goldman Sachs

• $350M for TensorWave, a Las Vegas-based all-AMD AI infrastructure provider, with funding from AMD Ventures, Magnetar Capital, and Maverick Silicon.

• $80M for PhoenixAI, a Menlo Park, CA-based agentic AI database platform, with funding from Sky9 Capital, Atypical Ventures, and Olive Technology Ventures.

• $60M for PointFive, a London and Tel Aviv-based AI infrastructure cost control startup, with funding from Accel, Salesforce Ventures, and Index Ventures.

• $40M for Forage, a San Francisco, CA-based EBT/SNAP payments platform, with funding from Mouro Capital, Nyca Partners, and PayPal Ventures.

Series C

• $435M for NewLimit, a South San Francisco, CA-based longevity therapeutics company, with funding from Founders Fund, Thrive Capital, and Lilly Ventures.

• $400M for NinjaOne, an Austin, TX-based IT management software platform, with funding from Sequoia Capital, ICONIQ, and Wellington Management.

• $300M for Mach Industries, a Huntington Beach, CA-based defense technology company, with funding from Infinite Capital, Ribbit Capital, Sequoia Capital, Bedrock Capital, and Khosla Ventures.

• $200M for Standard Bots, a Glen Cove, NY-based AI-native industrial robotics company, with funding from RoboStrategy and General Catalyst.

• $100M for ZutaCore, a Foster City, CA-based waterless data center cooling provider, with funding from Mitsubishi Electric, Carrier Ventures, and Samsung Venture Investment.

Series D

• $500M for Impulse Space, a Redondo Beach, CA-based in-space mobility company, with funding from 137 Ventures, Banner VC, and Founders Fund.

• $400M for Suno, a Cambridge, MA-based AI music-generation company, with funding from Bond Capital, IVP, and Forerunner.

• $290M for Isar Aerospace, a Munich-based launch services provider for small satellites, with funding from Molten Ventures, Island Green Capital, and HV Capital.

• $150M for Factorial, a Barcelona-based workforce management software company, with funding from General Catalyst, Atomico, and Four Rivers.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

SpaceX went public and AI filled the vacuum. The two most valuable private companies on Earth are now language-model shops. Anthropic doing nearly 2x OpenAI's revenue at a similar valuation tells us the market is pricing enterprise moats instead of hype. ByteDance at $186B revenue and a 3x multiple makes every AI name on this list look expensive.

How to Play It

The AI premium is real but fragile, if growth stalls, those 20x multiples compress fast, which makes ByteDance and Stripe the obvious value plays on secondaries. And if you want the biggest swing on the list, it's Waymo at 350x revenue - either autonomous driving scales or that $126B price tag corrects hard.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• Avidity Biosciences – acquired by Novartis for $12 billion

• D-Fend Solutions – acquired by Motorola Solutions for $1.5 billion

• Quantum Space – acquired by Inflection Point Acquisition for $1.2 billion

• Kumo AI – acquired by Nvidia for over $400 million

• Ona – acquired by OpenAI for an undisclosed amount

Top Venture-Backed IPOs

SpaceX (SPCX) | Hawthorne, CA-based aerospace and satellite communications company backed by Founders Fund, Sequoia, and Andreessen Horowitz. Raised $75B.

Quantinuum (QNT) | Broomfield, CO-based full-stack quantum computing company backed by Honeywell. Raised $1.68B.

Parabilis Medicines (PBLS) | Cambridge, MA-based biotech developing helicon peptide therapies for undruggable cancer targets. Raised $670M.

Liftoff Mobile (LFTO) | Redwood City, CA-based mobile ad-tech and performance marketing platform backed by Blackstone. Raised $437M.

Kardigan (KARD) | South San Francisco, CA-based clinical-stage cardiovascular precision therapeutics company. Raised $400M.

Deep Fission (FISN) | Berkeley, CA-based underground small modular nuclear reactor developer backed by Blue Owl. Raised $43M.

This Month at R136 Ventures

That’s all for this month. See you in June!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

Not as a VC, but as the guy running a 20,000-person technology organization inside one of Europe’s largest banks. I signed the checks and sat through procurement cycles, so I know how long it takes for new tool to get inside a Fortune 500 company.

The number I used to quote people for procurement was 16 months. 16 months from first meeting to a signed contract, and that was when things went well, but most of the time it never happened at all.

That’s the reason one number stopped me this spring…a16z found that 29% of the Fortune 500 and about 19% of the Global 2000 are now live, paying customers of at least one leading AI company.

To count, a company had to sign a top-down contract, convert the pilot, and put the product into real use. Roughly three years after ChatGPT launched. I watched this market for over two decades and I have not seen anything move that fast.

Everyone wants to know if AI is real inside the enterprise. For a third of the largest companies on earth, that question is closed. The better one is where it works, and why it works there but not down the hall. Look at that, and a pattern shows up. Once you see it, you can guess what comes next.

Where AI is currently winning: Code, support, and search

Three categories are pulling away from everything else, and they’re not the ones anybody picked in advance.

Code, support, and search.

Nobody decided these would win. They won because the work itself has a shape that fits what the models are good at right now.

Code is the clearest case. It’s text, has strict rules, and you can run it and see in a second whether it works. The feedback loop is tight. An engineer writes something, the model suggests something, the engineer runs it, and around it goes. No committee deciding if the output is good enough.

a16z puts productivity gain in the 10x to 20x range for top engineers. When the gain is that large on your most expensive people, nobody needs a long meeting about ROI.

At their developer conference, Dario Amodei disclosed 80x annualized growth in the first quarter of 2026, with the run rate crossing $30 billion, up from around $9 billion at the end of 2025.

Claude Code on its own passed $1 billion in annualized revenue within six months of launch. Discount those figures if you want (and there’s good reasons to be careful with run-rate math), but the direction is hard to miss. Coding tools have become part of how software gets built.

And that creates a problem most people are not thinking about. Compute. Once AI coding is standard for engineering teams, compute stops being something only the labs worry about and becomes an ordinary line item, the way cloud bills did fifteen years ago. Infrastructure turns into the constraint.

Thats why Amodei spent his stage time talking about a compute shortage instead of taking a victory lap.

Support sits on the other side of the building from engineering, and it works for the same reason. You can measure it. Tickets answered, resolution rate, satisfaction scores.

Support teams already run on scripts and escalation paths, which gives an AI agent a map to follow, and when it cannot handle something, the ticket goes to a human. That fallback is what makes it low risk to try. Low risk is what gets pilots approved.

Search is the quiet one, and probably the most universal. Every large company has the same buried problem. Information lives everywhere at once, in Slack and email and PDFs and the CRM and folders nobody has opened in years.

People lose hours just looking for things. AI search turns that mess into something usable. Not glamorous. But the pain is real and the payback is immediate.

Why legal and healthcare moved first

The industries adopting fastest reflect a similar story from another angle. Tech went first, which surprised nobody. But legal and healthcare following close behind is the part that surprised people, because both were famously hard markets for software.

Legal had long sales cycles and cautious buyers, and the work was too unstructured for old workflow tools to touch. What lawyers actually do all day is read dense text, reason over long documents, and draft and compare.

That’s exactly what these models are good at. So the value became obvious in a way it never was before. Harvey reportedly reached around $200 million in annualized revenue within three years of founding. That is a real business.

Healthcare runs the same way. The clinical work is complex, but a huge share of what happens in a hospital is administrative. Scribing, coding, prior authorizations, the byzantine rules for how care gets paid for.

Companies like Abridge and Ambience found they could take on those discrete chunks without forcing anyone to rip out their existing systems. That is the quiet key to enterprise adoption, and I learned it the hard way at the bank. Nobody wants to replace their system of record.

Everybody wants to pull labor out of a workflow that hurts. AI that sits on top of what already exists gets in the door fast. AI that asks you to tear something out waits a very long time.

The money is changing shape too

Once you see that AI works on measurable, repetitive, checkable work, you start to notice the same logic showing up where you would not expect it in how companies get bought.

Look at General Catalyst’s $6.3 billion take-private of American Express Global Business Travel, done through their Long Lake vehicle. Long Lake has already run dozens of acquisitions through its own AI platform.

The thesis is simple. If AI can change the cost structure of a service business, then an old service company with real distribution becomes a platform worth owning. You buy the customers, the workflow, the industry access, and then you put AI inside the operating model.

I have believed for a while that this is where the edge goes. Not to the AI-native startup with a clever wrapper, but to the boring incumbent that already has the customers and the proprietary data.

A thin layer on top of a model, with nobody paying for it, is going to struggle. A company with real customers, its own data, and the discipline to actually execute on AI becomes very hard to compete with. Distribution matters again. Domain trust matters again. Those things were starting to feel old-fashioned. They are not.

The org chart is getting thinner

The same force is showing up in headcount, which can be the uncomfortable part.

Coinbase reportedly cut around 14% of staff (roughly 700 people), with Brian Armstrong pointing to a return to an AI-native focus and the end of pure manager roles. Block made big cuts earlier and Pinterest and CrowdStrike are rethinking their numbers too. Some of this is just the cycle and ome of it is crypto winter, while some is ordinary cost discipline.

But AI gives management a new way to decide which work is worth keeping, and the most exposed work is the coordination layer. I mean the people who move information around without ever owning a hard decision, the ones who ask for a status update and then summarize it for someone higher up.

I have written before about the Jenga tower. AI pulls out the junior stabilizer pieces first, the rote work that quietly held the structure together, and what is left looks top-heavy and ready to tip unless you rebuild the org around it.

More people now have to own a workflow, a customer, a system, or some output you can point at. Coordination is hard to measure, and we already know what happens to work AI can measure better than we can.

Where the next wave forms

If you understand why code, support, and search went first, you can roughly call where this goes next, because the winners all share the same traits. The work is frequent and expensive, the output can be checked, the workflow is mostly digital, the buyer can see ROI quickly, a human can stay in the loop and the product helps the job before it tries to take the job.

Insurance fits almost perfectly, which is why Corgi reportedly hit a $1.3 billion valuation four months after its Series A, on a $160 million round led by TCV. Insurance runs on data, risk scoring, pricing, claims, and regulation, and AI can touch every link in that chain.

Accounts receivable fits too, which is why Fazeshift raised $22 million for autonomous AR. The work is structured, repetitive, and expensive to do by hand. Government fits earlier than most people expect, which is what Deepgram’s public sector push with Carahsoft and AWS signals.

Those workflows are massive, manual and full of voice and transcription and call centers. Once procurement opens, the category gets big fast, though it rewards patience, because security and auditability matter there more than speed ever will.

Fintech may be reopening on the same logic. The CLARITY Act is reportedly close to passage, with a July 4 target in the House, and if the final version keeps stablecoin rewards, it could be a real catalyst for digital assets infrastructure, with Circle reportedly up around 20% and Bitcoin above $82K on the news.

For anyone building here, the price of crypto was never the real question. I do not invest in crypto for the speculation. We backed eToro and Blockchain.com because I saw them as infrastructure.

The question has always been whether digital assets can move from speculation into regulated financial plumbing, the stablecoins and payments and custody and cross-border flows.

All of that gets more useful the moment the rules get clearer. And clearer rules are not a threat to the serious builders. If a license is hard for you to get, it is hard for your competitors too.

What I’m watching

So the useful question now is simple. Where does AI turn messy labor into output you can measure? THAT’S where the adoption is. Not everywhere..not evenly, but in the specific places where the work, the tech, and the buyer’s incentive all line up.

Some markets will move slower. Physical work is harder…relationship work is harder. Regulated work stays hard until the trust layer is built.

The capabilities keep climbing while we wait. a16z’s read of the GDPval benchmark shows accounting and auditing up nearly 20%, and detective work close to 30% in just four months. Jumps like that tend to produce new products and new companies in the domains right behind them.

The mistake people keep making is judging enterprise AI by the average pilot success rate, because the average hides the movement. In some departments AI is still a slide in a strategy deck.

In others, it is already in the budget, in the workflow, and in the headcount plan. That gap is where the whole opportunity sits. AI found its first real work. The market follows the work…it always has.

What You’ll Learn:

• The early days of conversational AI before transformers

• Building emotional intelligence into software

• Why Replika had 1.5M users on the waitlist before launch

• The rise of ChatGPT and OpenAI

• Why emotional AI may become bigger than assistants

• The future of AI relationships and digital companions

• Why AI could replace millions of jobs

• The next generation of AI interfaces through Wabi

• Consumer AI vs enterprise AI

• The future battle between Google, OpenAI, Anthropic, and startups
  

This conversation explores one of the biggest questions of our time: can AI truly understand humans, and what happens if it does?

The Operator’s Lens

Welcome to May’s growth newsletter, where we track what matters in late-stage venture.

What's interesting is who got through the IPO window. Cerebras, Fervo...both traded up 40% on day one, both infrastructure plays that need years to scale. Market is pricing inevitability. If you're building software that improves margins by 10%, you're still fundraising in 2021 terms. If you're building energy or silicon that reshapes cost curves for next decade, you're getting 100x multiples and clearing exits. Maybe that's same thing as saying only hard problems get liquidity now.

Let’s get into it.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $300M for Amca, an El Segundo, CA-based enterprise technology company, with funding from Caffeinated Capital, Lightspeed Venture Partners, and Andreessen Horowitz.

• $275M for Cowboy Space Corporation, a San Carlos, CA-based orbital infrastructure builder, with funding from Index Ventures, IVP, and Blossom Capital.

• $230M for Armada, a San Francisco, CA-based modular AI data center company, with funding from Overmatch, BlackRock, and 8090 Industries.

• $200M for Blitzy, a Cambridge, MA-based autonomous software development platform, with funding from Northzone.

• $170M for Radar, a New York, NY-based RFID-powered inventory platform, with funding from Gideon Strategic Partners, Nimble Partners, and Align Ventures.

Series C

• $355M for Modal Labs, a New York, NY-based AI infrastructure startup, with funding from Redpoint Ventures, General Catalyst, Accel, and Menlo Ventures.

• $300M for Decart, a San Fran and Tel Aviv-based generative AI frontier lab, with funding from Radical Ventures, Sequoia Capital, Nvidia, Adobe Ventures, and Andrej Karpathy and others.

• $250M for Exa Labs, a San Francisco, CA-based AI search tools company, with funding from Andreessen Horowitz.

• $150M for Oishii, a Jersey City, NJ-based indoor vertical farming company, with funding from SPARX Asset Management, Nomura Real Estate Development, and MISUMI Group.

• $100M for Nourish, a New York, NY-based metabolic health clinic, with funding from Menlo Ventures, Thrive Capital, and Index Ventures.

Series D

• $550M for ElevenLabs, a New York, NY-based AI voice and audio technology company, with funding from BlackRock and NVIDIA.

• $200M for Mercury, a San Francisco, CA-based banking platform for startups, with funding from TCV, Andreessen Horowitz, Coatue, and CRV.

• $150M for Farther, a New York, NY-based AI-integrated wealth management platform for high-net-worth clients, with funding from General Atlantic, Bessemer, MassMutual Ventures and others.

• $120M for Elliptic, a London, UK-based crypto compliance and analytics company, with funding from One Peak Partners with participation from Nasdaq Ventures and Deutsche Bank.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

SpaceX surged another 40% to $1.75T - now worth more than the next three combined, driven by Starlink’s commercial traction and the xAI vertical integration narrative fully pricing in. The middle compressed, Databricks flat and Stripe slightly upwhile ByteDance and Ant Group anchor the bottom with 3x multiples, reminding us that not everything trades like AI.

How to Play It

SpaceX at $1.75T is the obvious winner but hardest to access; Databricks at $134B remains the clearest entry with real revenue and institutional liquidity. Waymo's 315x multiple screams bubble, but if robotaxis scale, it's the cheapest exposure to post-Uber mobility. Stripe and xAI are the underpriced sleepers: Stripe's AI-native payments infrastructure is still trading at 32x (half of xAI's 66x), and xAI at $250B standalone is a bet on Grok eating search share from Google.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• Stainless - acquired by Anthropic for an undisclosed amount

• Weights.gg - acquired by OpenAI for an undisclosed amount

• Symmetry Systems - acquired by Zscaler for an undisclosed amount

• ARI (Assured Robot Intelligence) - acquired by Meta for an undisclosed amount

• Jit - acquired by Torq for an undisclosed amount

Top Venture-Backed IPOs

$6.85B raised across 2 major venture-backed tech IPOs in May 2026. Cerebras' $5.55B debut was one of the largest AI infrastructure IPOs on record and a direct signal that the compute buildout trade is very much alive.

• Cerebras Systems (CBRS) | Sunnyvale, CA-based AI chip maker with wafer-scale processors built for large-scale inference. Raised $5.55B.

• Fervo Energy (FRVO) | Houston, TX-based enhanced geothermal systems developer backed by J.P. Morgan and BofA. Raised $1.3B.

This Month at R136 Ventures

That’s all for this month. See you in June!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

The whole model assumed a simple question: who is coming in, and what are they carrying?

That model worked because the threats were mostly external. Someone trying to break through the wall, malware hiding in a software package, a compromised credential being used to sneak past the gate.

You built higher walls, better scanners, smarter detection systems and for a long time that was enough. But something fundamental changed in the last two years, and most companies are not ready for it.

The change is AI agents.

Not the kind that retrieve information or answer questions, but the kind that make decisions. They access databases, call APIs, trigger workflows, move money. They operate inside your infrastructure with real authority, and they are spreading fast.

McKinsey reported that 78% of organizations now use AI in at least one business function (up from 55% in 2023). More than half of enterprises have actively deployed AI agents, and 39% of those are running more than ten agents in production.

Gartner projects that by 2028, a third of all enterprise software will incorporate agentic AI, which is more than thirty-fold increase from less than 1% in 2024.

I ran technology for one of the largest digital banks in Europe for six years. We scaled from zero to 45 million monthly active users. At that scale, security is not abstract; it is operational.

Every new capability you give your systems, every integration you open or API you expose creates something that somebody will try to exploit. We understood that because we were dealing with human actors, humans you can identify, authenticate, monitor, and hold accountable. But AI agents are different.

They are self-programmable, can be prompted, manipulated, redirected. In airport metaphor, imagine that one of the authorities inside the perimeter can be convinced to become a criminal, or simply acquires permissions it was never supposed to have.

That is how AI agents work inside enterprise systems today, and the security stack we built over the past 20 years was not designed for autonomous actors operating inside your own walls.

The Threat Sitting Inside Your Firewall

Headlines are now full of stories about hackers using AI, and that threat is real. Over 80% of cyberattacks now utilize AI. AI-generated phishing has driven a 703% surge in credential phishing attacks.

Deepfake incidents rose over 60% last year, with three-quarters of those fakes designed to impersonate CEOs and C-suite executives. The global average cost of a data breach hit $4.88 million, a record. All of that is serious, but it is also the obvious part.

The less obvious threat is one sitting inside your firewall, which is your own AI agents. You gave them the keys, you gave them access to customer data, financial systems, internal tools.

The question most companies cannot answer right now is simple: do you actually know what your agents are doing? Can you monitor their behavior in real time, can you catch when an agent starts drifting from its intended purpose, or when someone figures out how to make it do something it was never supposed to do?

OWASP identified prompt injection as a top security risk for large language models. This is exactly what it sounds like: someone crafts an input that causes your AI to ignore its instructions and follow new ones.

It does not require breaking into your network because the agent is already inside, the attack surface is the agent itself. And it gets worse at scale. Most enterprises are not running one agent, they are running dozens.

Less than 10% of organizations have successfully scaled AI agents in even a single business function, which means the majority of these deployments are in a messy, semi-operational state. Partially integrated, partially monitored, partially understood.

60% of IT professionals believe their organizations are not prepared to counter AI-generated threats. Only 18% of security teams have fully adopted AI-powered cybersecurity tools. These are not small numbers, they describe a systemic gap between deployment speed and security readiness

Why This Isn’t Getting Fixed

The core problem is incentives. Every company is racing to deploy AI, speed is rewarded. The board wants to hear about AI strategy, the market rewards companies that ship AI-powered features. Nobody gets promoted for slowing things down to think about security architecture.

I see this from the investor side. We look at hundreds of companies building AI applications, the energy is extraordinary. But the pattern is familiar: deploy first, secure later.

We saw this with cloud computing a decade ago, we saw it with mobile. The difference is that AI agents have more autonomy, more access, and more ability to cause damage than any previous technology deployed inside the enterprise.

The data confirms the mismatch. Between January 2024 and December 2025, startups in AI security raised a total of $8.5 billion. That sounds like a lot until you look closer: only $414 million of that, less than 5%, went to companies building solutions specifically to secure AI models, LLMs, and agentic systems.

The rest went to traditional cybersecurity tools that happen to use AI for defense, things like network security, threat detection, SOC automation. Important work, but not the work of securing the AI systems themselves.

Meanwhile, 93% of organizations say they understand AI risks. That sounds reassuring until you learn that fewer than half have an AI governance framework in place, only 43% have an AI-specific incident response plan.

75% of technology leaders cite governance as primary concern when deploying agentic AI, yet a third of organizations admit they lack a clear strategy for implementing it securely.

This is one of the reasons we spend so much time studying AI security operations at R136. Not because it is trendy, but because the gap between how fast agents are being deployed and how slowly security is catching up is one of the biggest mismatches I have seen in twenty years of technology investing.

What the Winners Are Building

The companies that will define this space share a few characteristics I keep seeing. First, they are building prevention-first architecture, not just better detection.

The traditional model waits for something bad to happen and then responds, but with AI agents operating autonomously at speed, by the time you detect a problem the damage may already be done. You need systems that constrain agent behavior before it goes wrong.

Second, deep observability. Not just logging what an agent did after the fact, but understanding what it is doing in real time: what permissions is it using, what data is it accessing, is its behavior consistent with its intended function? This is closer to how you would monitor a new employee with broad access than how you would monitor a piece of software.

Third, governance that treats AI agents more like employees than tools. This sounds strange, but think about it: when you hire someone and give them access to sensitive systems, you have an onboarding process, you have access controls, you have periodic reviews, you have the ability to revoke access.

Most companies have nothing equivalent for their AI agents, they deploy them with broad permissions and hope for the best.

The market is beginning to price this as a massive opportunity. The AI cybersecurity market is projected to grow from roughly $25 billion in 2024 to somewhere between $65 billion and $94 billion by 2030. The AI governance market alone is expected to reach nearly $6 billion by 2029, growing at 45% annually.

Venture funding is flowing in, companies like Noma Security ($100M), Lakera, and Prompt Security are raising meaningful rounds to build specifically for this problem. I think the market might be underestimating the size of this.

The projections assume a relatively orderly adoption curve, they do not fully account for what happens when something goes seriously wrong at a major company because of an unsecured AI agent. That event will reprice the entire market overnight.

The Risk Nobody Wants to Discuss

I have been investing in technology for twenty years, and before that I spent over decade building it at large institutions. If there is one pattern I have learned to recognize, it is this: the biggest risks are the ones nobody wants to discuss.

Not because they are invisible, but because discussing them requires admitting something uncomfortable, and admitting it might slow things down.

AI security is exactly that kind of risk. Talking about it means acknowledging that the AI systems companies are rushing to deploy might themselves be the vulnerability.

That is not comfortable conversation for a CEO who just told the board that AI is the company’s strategic priority, it is not a comfortable conversation for a CTO who has spent the last year building an AI platform. And it is definitely not a comfortable conversation for investors who are betting on the AI application wave.

But comfortable conversations are not where the important insights live. The companies that take AI security seriously now, that build governance and observability and prevention into their agent architectures from the start, will still be standing when something goes wrong. And something will go wrong. The only question is whether you are ready for it.

💡 What you’ll learn:

• Why selling directly to SMBs failed — and what worked instead

• The real reason agencies became the growth engine

• How a simple pivot changed everything

• Why HighLevel never relied on a traditional sales team

• The strategy behind 40% lifetime affiliate commissions

• How to scale a SaaS company profitably from day one

• The truth about product-market fit most founders misunderstand

• How AI is changing (but not replacing) service businesses
  

The Operator’s Lens

Welcome to April’s growth newsletter, where we track what matters in late-stage venture.

April showed where the money is: nuclear reactors, obesity drugs, proteomics, and AI chips. The late-stage market hasn't frozen, but it's just betting on molecules, atoms, and silicon. If you're building software or marketplaces, you're in a different fundraising environment.

Let’s get into it.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $250M for TENEX.AI, Sarasota, FL-based AI-native MDR platform, with funding from Crosspoint Capital Partners, Shield Capital, and DTCP.

• $137M for Sidewinder Therapeutics, San Diego-based biotech company, with funding from Frazier Life Sciences, Novartis Venture Fund, and OrbiMed.

• $125M for Ray Therapeutics, Berkeley, CA-based biopharma, with funding from Janus Henderson Investors, Adage Capital Management, and Franklin Templeton.

• $100M for Neomorph, San Diego, CA-based therapeutics developer, with funding from Deerfield Management, Regeneron Ventures, and Longwood Fund.

• $38M for Luminai, San Francisco-based healthcare AI platform with funding from Peak XV Partners, Define Ventures, and General Catalyst.

Series C

• $650M for Slate Auto, Troy, MI-based electric pickup truck maker, with funding from TWG Global, General Catalyst, and Slauson & Co.

• $170M for Glydways, San Francisco, CA-based autonomous pod transit startup, with funding from Suzuki Motor Corporation, ACS Group, and Khosla Ventures.

• $150M for Factory, San Francisco, CA-based AI coding platform, with funding from Khosla Ventures, Sequoia Capital, and Insight Partners.

• $140.8M for TeraWatt Technology, San Francisco-based battery developer, with funding from Khosla Ventures, Temasek, JIC Venture Growth Investments, and others.

• $120M for Omni, San Francisco, CA-based AI analytics platform, with funding from ICONIQ, Theory Ventures, and First Round Capital.

• $90M for Coder, Austin, TX-based AI infrastructure platform, with funding from KKR, Qube Research & Technologies, Uncork Capital, and others

Series D

• $1.75B to Saronic Technologies, Austin, TX-based autonomous vessel company with funding from Kleiner Perkins, Advent International, and Bessemer Venture Partners.

• $693M for Tenstorrent, Austin, TX-based AI chip and computing solutions company, with funding from Samsung, Bezos Expeditions, and AFW Partners.

• $150M for Matillion, a Denver, CO-based cloud-native data transformation platform, with funding from General Atlantic.

• $70 million to Censys, to MI-based defense-tech platform with funding from Morgan Stanley Expansion Capital with participation from Decibel Partners and Greylock Partners.

• $50M for Ushur, a Santa Clara, CA-based customer experience automation platform for regulated industries, with funding from Third Point Ventures.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

SpaceX just hit $1.25 trillion (56% jump in one month) becoming the first private company to crack a trillion. That’s Starlink revenue hitting scale and xAI creating a vertical monopoly from chips to satellites. OpenAI crept to $852B, proving secondary buyers still see room despite the conditional $110B structure. Anthropic held at $380B (Pentagon rejection intact).

The middle froze. Databricks raised $5B without moving its valuation. Stripe took $14.6M (likely insiders). Reliance Retail sneaking into the top 10 at $101B is the outlier.

How to Play It

SpaceX at $1.25T is the safest asymmetric bet if you can find shares. Databricks remains the cleanest entry: $5B raise, flat valuation, real revenue. OpenAI's secondary premium is compressing. Stripe's AI payments play is still underpriced.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• Kelonia Therapuetics — acquired by Eli Lilly Company for $7B.

• TBPN — acquired by OpenAI for $100M+

• Grüns — acquired by Unilever for $1.2B

• SuperNODE (by GigalO) — acquired by d-Matrix for an undisclosed amount

• Zengo — acquired by eToro for $70M

Top Venture-Backed IPOs

$1.84B raised across 3 venture-backed tech IPOs so far in April 2026. X-Energy's $1B+ raise was the standout, marking one of the largest nuclear tech IPOs

• X-Energy (XE) | Rockville, MD-based nuclear reactor and TRISO fuel developer backed by Amazon. Raised $1.02B in IPO.

• Alamar Biosciences (ALMR) | Fremont, CA-based precision proteomics platform for early disease detection. Raised $191M in IPO.

• Kailera Therapeutics (KLRA) | Waltham, MA-based obesity drug developer with GLP-1 therapies. Raised $625M in IPO.

This Month at R136 Ventures

That’s all for this month. See you in April!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

In three weeks, those four people built him a full Notion replacement. Custom, internal, deployed company-wide.

Now they’re building a CRM to replace his current licensed solution, with ERP next in line, then HR platform, then business process management system. His vibe coders are systematically dismantling his entire SaaS stack, one tool at a time.

But the software replacement is only half of story.

He replaced seven M&A analysts with two people. These aren’t traditional analysts with finance degrees and Bloomberg terminals. They’re AI-native operators who build agents, train models on historical deal files, and produce faster, more consistent output than the team of seven ever did.

In HR, he’s planning to go from 37 people to 7. Across the whole company, he believes he can take his non-technical workforce from nearly a thousand to roughly one hundred, by end of 2026.

He’s essentially restructuring entire logic of how a company operates.

The Week the Market Finally Noticed

In early February, Anthropic released a set of industry-specific plugins for its Claude Cowork platform, covering legal, sales, finance, and data analytics. Within 48 hours, $285 billion in market value evaporated from software, financial services, and data companies across three continents.

Thomson Reuters fell 18% in single session. RELX (parent of LexisNexis), dropped 14%. Wolters Kluwer shed 13%. Analysts coined a term for it..the SaaSpocalypse.

Two days later, Anthropic released Claude Opus 4.6, a model that can spin up and coordinate entire teams of AI agents autonomously. FactSet dropped over 9%, and S&P Global, Moody’s, and Nasdaq followed. The bleeding spread from legal to financial services to asset management.

The market wasn’t reacting to a product. It was reacting to an inevitability.

I’ve spent 20 years in technology. I ran a digital transformation with 20,000 engineers and spent billions on enterprise software. I know exactly how Fortune 500 companies buy, implement, and get locked into SaaS contracts. That world is ending.

The question I keep turning over is simple: what comes next?

Why This Is Happening Now

Three forces hit at the same time and together they created something much bigger than any one alone.

The model layer moved into the application layer. When Anthropic launched Cowork plugins, it didn’t just offer a better chatbot. It also moved from selling picks and shovels to mining the gold itself.

When the foundation model maker starts shipping ready-made vertical solutions, the platform itself becomes the competitor. Every SaaS vendor sitting on top of that platform suddenly has to ask a very uncomfortable question about their own future.

Vibe coding hit escape velocity. Tools like Claude Code, Cursor, and Replit have made it possible for non-engineers to build production-grade internal software in days, not quarters. The skill isn’t writing code anymore, it’s describing what you need and iterating on what the model produces.

A Facebook group of enthusiasts can now out-ship an enterprise vendor’s product team. Think about what that means for every company paying millions a year for tools they can’t customize and don’t own.

The cost calculus flipped. When your Notion subscription costs $15 per seat per month across 5,000 employees, that’s $900,000 a year for a tool you can’t customize and now it’s replaceable in three weeks. Multiply that across every SaaS line item on the P&L, and CFOs start asking very uncomfortable questions.

Service as a Software: The Inversion

For twenty years, the dominant model was Software as a Service. You buy platform, pay per seat, and the vendor handles the infrastructure. SaaS was genuinely revolutionary because it eliminated capital expenditure of on-premise software.

What’s emerging now is the exact inversion.

You don’t buy a platform…you buy a capability. You hire vibe coders, or deploy AI agents, who deliver software as the output of a service engagement.

The software they produce is custom, internal, disposable, and perpetually evolving. There’s no vendor lock-in because there’s no vendor. There’s no per-seat pricing because you own the code.

Same acronym. Reversed polarity.

The old SaaS: we built the software, you subscribe to the service.

The new SaaS: you buy the service of building, and the software is what you get.

I know this sounds like a semantic trick, but the economic implications are enormous. The entire SaaS business model depends on recurring revenue from customers who can’t easily leave. When the switching cost drops to near zero and the build cost drops right alongside it, that recurring revenue becomes very fragile.

Does It Scale?

My friend’s company has 5,000 employees. He’s in gaming, an industry already native to rapid iteration. You could argue he’s an edge case.

But does this model work at 50,000 employees? Or 300,000?

I think it does, and reason is straightforward: the constraint was never vision. It was tooling.

When building custom internal software required a 50-person engineering team, an 18-month roadmap, and a $10 million budget, it made perfect sense to subscribe to Salesforce or Workday or ServiceNow. That tradeoff has now collapsed. The four vibe coders my friend hired cost a fraction of a single enterprise software contract.

As AI agents become more capable, even the vibe coders become optional for certain workflows. The agents themselves become the builders.

We’re heading toward a world where the largest enterprises don’t have software stacks in the traditional sense. They have AI-native operating systems that continuously generate, test, and deploy internal tools based on the company’s actual workflows.

Maybe that sounds like science fiction. But three weeks to replace Notion with four people from a Facebook group would have sounded like science fiction two years ago.

What Dies and What Survives

Not all software is equally exposed. Systems of record with deep regulatory moats... core banking, medical records, air traffic control... those won’t be vibe-coded away next quarter.

But the layers above those systems, the analytics dashboards, the workflow automations, the project management tools, the CRM interfaces... those are now in what I’d call the kill zone.

Piper Sandler analyst Billy Fitzsimmons warned that “seat compression and vibe coding narratives could set a ceiling on multiples.” He was being polite. If companies can build their own tools and slash headcount simultaneously, the entire pricing model of enterprise software breaks.

What survives? The model providers. The compute layer. The companies that own the intelligence, not the interface. And a new class of AI-native operators... the vibe coders and agent builders who sit between the model and the business problem.

I’ve been thinking about this through the lens of my own experience running massive technology teams. The 16-month enterprise procurement cycle I used to navigate existed because building software was hard and slow. When building becomes fast and cheap, that process compresses. And the vendors who depended on that friction... they lose their moat.

Maybe most surprising thing about this moment is how fast it moved from theory to reality. A set of plugins in late January. A model upgrade in early February. And $285 billion gone in 48 hours.

The frontier isn’t gone. It just looks different now.

The Operator’s Lens

Hey folks, welcome to March’s newsletter, where we track what matters in late-stage venture.

March was the month the market started asking for receipts. OpenAI closed the largest private round in history but look at the structure: compute credits, conditional tranches, milestones tied to an IPO.

Even at $840B, investors aren’t writing blank checks anymore.

After months of repricing and mega-rounds, March was a pause not a pullback, but a market catching its breath.

IPOs stayed narrow, the window is open, but only for the obvious winners.

Late-stage capital is getting disciplined. The days of “growth at any price” are behind us. Now it’s about structure, milestones, and proof—not just promise.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $500M for Nexthop AI, a Santa Clara, CA-based artificial intelligence company, with funding from Lightspeed Venture Partners, Andreessen Horowitz, and Altimeter.

• $385M for Genspark.ai, a Palo Alto, CA-based AI workspace platform, with funding from Emergence Capital, SBI, and Mirae Asset.

• $165M for Sunday, a Mountain View, CA-based household robot maker, with funding from Coatue, Tiger Global, and Benchmark.

• $155M for Nominal, Los Angeles based data platform with funding from Founders Fund, Sequoia, and General Catalyst.

• $80M for Juicebox, a San Francisco-based AI-native recruiting platform, with funding from DST Global, Sequoia Capital, and Coatue.

Series C

• $230M for Science Corporation, Alameda based with funding from Lightspeed Venture partners, Khosl Ventures, Y Combinator and others.

• $100M for ORO Labs, a Palo Alto, CA-based procurement orchestration platform, with funding from Brighton Park Capital, Goldman Sachs, Norwest Venture Partners, and others,

• $100M for Cape, an Arlington, VA-based privacy-first mobile carrier, with funding from Bain Capital Ventures, Goldman Sachs, 01 Advisors, 137 Ventures, Definition, and others.

• $65M for Sage, a New York-based AI platform for aging care, with funding from Growth Equity at Goldman Sachs Alternatives, IVP, and Goldcrest Capital,

• $40M for Turquoise Health, a San Diego-based healthcare pricing and payments platform, with funding from Goldman Sachs, Andreessen Horowitz, and Adams Street Partners.

Series D

• $400M for Replit, San Franscisco based AI coding platform with funding from Georgian Partners, Andressen Horowitz, Accenture Ventures, Prysm Capital and others.

• $300M for Verily, TX based healthcare data platform with funding from Series X Capital, Alphabet, UCHealth and others.

• $250M for Findhelp, Texas based social care platform with funding from The Rise Fund.

• $150M for Grow Therapy, a New York-based mental health therapy platform, with funding from TCV, Goldman Sachs, and Sequoia Capital.

• $143M for Frore Systems, a San Jose-based cooling for AI chips company, with funding from MVP Ventures, Fidelity Management & Research Company, Top Tier, and others.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

OpenAI’s latest $110B round looks massive until you realize most of it is compute credits and conditional tranches, not cash. At $840B, they’re priced for perfection while burning $35B+ annually. Stripe quietly launching AI agent payment rails is the sleeper: if machines start moving money, they just became the tollbooth.

How to Play It

OpenAI at $840B valuation is priced for perfection, any stumble and you're holding the bag. Databricks ($134B, $5.4B revenue, flat valuation) is the only clean entry. Stripe's AI payments play is underpriced. SpaceX IPO whispers at $1.75T make $800B look like a floor.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• Promptfoo — acquired by OpenAI for an undisclosed amount

• Forethought — acquired by Zendesk for an undisclosed amount

• RIVR — acquired by Amazon for an undisclosed amount

• Cal AI — acquired by MyFitnessPal for an undisclosed amount

• Moltbook — acquired by Meta for an undisclosed amount

Top Venture-Backed IPOs

The US IPO market in March saw minimal activity with only a handful of small deals, signaling that the IPO window remains selective. Large-cap names are still waiting for clearer macro signals before testing public markets.

• Swarmer (SWMR) | Austin-based AI-powered autonomous drone defense systems company. Raised $15M in IPO.

This Month at R136 Ventures

Ventures from the Valley Episode #25: The Platform Behind $1B in Startup Raises with Stephane Nasser

He's helped 25,000 founders raise $1B+ without taking a dime of equity. This month, Victor sits down with Stephane Nasser of OpenVC to unpack what actually moves investors (and it's not your slides).

Listen here

That’s all for this month. See you in April!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

For most of the last decade, voice AI was something you tolerated. You called a bank, got an IVR system, said “billing,” in as clear a voice as you could and somehow you still ended up in the wrong queue.

The technology was clunky and the gap between what people expected and what they got was embarrassing…That gap is closing super fast.

Data from 400 surveyed business leaders in the US highlights this:

1. 80% of organizations use traditional voice agent systems, yet only 21% are very satisfied

2. 67% of organizations consider voice AI core to their product and business strategy

3. 15% of organizations are already actively developing voice AI agents and 98% of these plan to deploy within 12 months

What’s actually now changing isn’t that voice AI got better (though it has). It’s that the cost of failure got a lot higher. Enterprises are deploying AI agents at scale now.

Those agents don’t live in chat windows anymore but in phone calls, customer service lines, sales conversations, clinical documentation workflows. If the voice layer breaks, the whole thing breaks and “slightly better” isn’t good enough when processing millions of conversations a day.

Legacy voice systems were built on heuristics which made them rule-based and brittle. They worked fine when the use case was narrow but fail the min you need real-time, contextual, full-duplex conversation at enterprise scale. That’s the gap that opened up, and it’s the gap that serious infrastructure companies are now racing to fill.

This is why we invested in Deepgram.

Why Deepgram?

Founders Scott Stephenson and Adam Sypniewski started Deepgram in 2015 while working two miles underground in a physics lab. They were analyzing massive audio waveforms, looking for tiny signals buried in noise, and needed speech recognition software to do it.

Everything available at the time was inadequate..so they built their own.

What made that decision consequential wasn’t just that they built..it’s how they built it. They didn’t iterate on what already existed.

They threw out the heuristic-based approaches entirely and built a full end-to-end deep learning architecture from scratch, under constraint, early. That choice compounded over time in ways that are hard to replicate now.

Today they own the full stack: data labeling, model training, inference, and the application layer. That vertical integration is where the performance numbers come from. Nova-2, their speech-to-text model, and Aura-2, text-to-speech model, are the fastest in their categories by independent benchmarks.

Aura-2 is designed specifically for the economics of enterprise deployment. And Flux, their newest model, solves something nobody else has cleanly handled yet: interruptions.

In real conversations, people talk over each other and legacy systems break when that happens. Flux was built for exactly that problem.

The customer list reflects where they’ve gotten to. NASA is now using it for space-to-ground communications. Five9 runs billions of call minutes through the platform. Twilio embedded it as their core voice infrastructure.

Over 1,300 organizations and 200,000 developers build on Deepgram’s APIs.

And in February 2026, IBM named Deepgram their first voice AI partner, integrating it directly into watsonx Orchestrate for clients in healthcare, finance, and government.

That IBM deal matters more than it might look on the surface. This not just a pilot or a partnership announcement. IBM is making Deepgram the voice layer of their entire enterprise AI product.

Scott Stephenson has said publicly that his goal is to pass the “Audio Turing Test” at scale in 2026, meaning AI voice that’s indistinguishable from human in real enterprise conversations.

Whether they hit that or not, the ambition tells you something about how they think about the problem they’re solving.

How We Looked at It

The first thing we pushed on with Deepgram was the competitive moat. Google has voice AI, and so does Amazon and Microsoft. All three have unlimited compute, massive distribution, and the ability to price very aggressively.

So the question we kept coming back to was why doesn’t one of them just win this?

Deepgram has been compounding in a very specific direction for over a decade, and that kind of lead is harder to close than it looks. They built on end-to-end deep learning from day one, while Google and Amazon started from NLP-based foundations.

They co-designed their models and physical infrastructure together since the beginning, which gives them hardware-level optimizations you can’t easily retrofit onto a system built differently. And voice AI has been their only product since 2015.

Google and Amazon are managing hundreds of products across dozens of categories. Deepgram has been solving one problem, at depth, for ten years. That focus shows up in the product and in the go-to-market.

Years of building specifically for enterprise deployments means the product reflects what those customers actually need, not what a general-purpose platform can accommodate on the side.

The only thing we were skeptical about was whether any company could maintain a meaningful speed and accuracy advantage over hyperscalers for more than 18 to 24 months.

Then what changed our mind was the customer evidence.

When you talk to enterprises that have been running millions of call minutes through Deepgram and they explain what switching would actually involve, the acoustic models trained on their specific data, the on-premise configurations, the customization built up over time, you see that the moat isn’t purely technical.

It compounds through the relationship and deepens with every deployment.

We also looked really hard at the founder thesis. Two particle physicists who built voice AI from scratch, underground, while analyzing waveform data for dark matter experiments. That’s not a resume line. That’s evidence of a specific way of thinking about problems.

Why It Fits R136’s Thesis

We’ve said for a while now that the foundational model race is largely over. The interesting question now is what gets built on top of all that, and who builds the infrastructure that makes it actually usable inside real businesses.

Deepgram fits that framing exactly. They’re not trying to build a better LLM, instead they’re the API layer that turns voice AI from a demo into something enterprises can deploy at scale and trust with millions of real conversations.

Think about what Stripe did for payments. Allowed developers to stop worrying about payment rails and to just build. That’s what Deepgram is doing for voice.

It also sits naturally alongside the rest of the portfolio. Airwallex, Blockchain.com, now Deepgram. Companies building essential infrastructure in large markets, with real customers and real switching costs, and genuine reasons to still be around a decade from now.

The upside is simple to say and harder to guarantee. If AI agents become the norm over the next five years and voice becomes how people actually interact with them, whoever owns that infrastructure layer is in a very good position.

The $130M round at a $1.3B valuation reflects where Deepgram is today. How right we are of course depends mostly on how big the category gets.

Closing Thought

I’ve been in technology long enough to know that the interface layer matters more than most people expect.

The decisions that actually moved retention were almost always about how people interacted with the product, not what the product could technically do.

Voice is the most natural human interface that exists and was always going to matter. What changed is that the infrastructure finally caught up to what’s actually required.

That’s what we’re backing with Deepgram. Not a bet on voice AI as a trend but a bet on a specific team that spent a decade building the right architecture, at the right time, for a real market need.

Whether we’re right, we’ll find out. But the evidence is there, and the team has earned the benefit of the doubt.

Congrats to Scott, Adam, and the entire Deepgram team on the milestone!

In this episode, I sit down with Stephane Nasser, Co-Founder & CEO of OpenVC - a free fundraising platform connecting 25,000+ founders with 6,000+ investors, with over $1B in funding facilitated since 2021.

Stephane has reviewed thousands of pitch decks and seen just about every mistake. His take is blunt..most founders aren’t lying, they just don’t know how to speak investor language.

We get into why investors respond to scarcity more than slides, why the best deals never hit a deck at all, and why Stephane built a platform that deliberately hides founder names from investors.

We cover:

• Why FOMO moves deals faster than fundamentals

• The pitch deck mistakes that kill deals before the first meeting

• How OpenVC’s blind deal flow feature gets 95%+ open rates

• Why warm intros still dominate, and how to engineer them

• What “climbing the learning curve” actually means before you pitch

• How storytelling is moving from decks to social media

• Whether venture capital itself could be disrupted

If you’re raising Pre-Seed, Seed, or Series A, or sourcing early-stage deal flow, this one’s worth your time.

The Operator’s Lens

Hey folks, welcome to February’s newsletter, where we track what matters in late-stage venture.

The market isn't rising, it's splitting. February's rounds made that cleaner: tier-1 capital went to hard infrastructure and AI with real moats, while everything else competed on fundamentals and hope.

The revenue bar to raise has never been higher and graduation rates have never been lower. If your story doesn't anchor to one of those winning verticals, your Series C conversation is going to be harder than your Series B was.

Top 5 Tweets

The smartest takes on tech and venture capital this month, pulled from X chaos.

Top 5 Growth-Stage Rounds

Here's a snapshot of where the late-stage VC money went the past month:

Series B

• $270M for San Francisco-based Bedrock Robotics with funding from CapitalG, Valor Atreides AI Fund, 8VC, and others.

• $230M for Reno-based Positron with funding from ARENA Private Wealth, Jump Trading, and Unless.

• $150M for San Francisco-based Goodfire with funding from B Capital, Menlo Ventures, Lightspeed Venture Partners, and others.

• $140M for Scotts Valley, CA-based Heron Power with funding from Andreessen Horowitz, Breakthrough Energy Ventures, Capricorn Investment Group, and others.

• $125M for Boston-based Code Metal with funding from Salesforce Ventures, Accel, B Capital, and others

Series C

• $300M for El Segundo, CA-based Skyryse with funding from Autopilot Ventures, Fidelity Management & Research Company, ArrowMark Partners, and others.

• $125M for San Francisco-based Kindred with funding from NEA, Index Ventures, Dylan Field, and others.

• $100M for San Francisco-based Render with funding from Georgian, Addition, Bessemer Venture Partners, General Catalyst, and others.

• $70M for San Francisco-based TRM Labs with funding from Blockchain Capital, Goldman Sachs, Bessemer Venture Partners, and others.

• $50M for New York-based Pepper with funding from Lead Edge Capital, ICONIQ, Index Ventures, Greylock, and others.

Series D

• $300M for Bellevue, WA-based Temporal with funding from Andreessen Horowitz, Lightspeed Venture Partners, Sapphire Ventures, and others.

• $232M for Mountain View, CA-based Lunar Energy with funding from Activate Capital, B Capital, Prelude Ventures, and others.

• $210M for NYC-based Talkiatry with funding from Andreessen Horowitz, Insight Partners, OMERS Ventures, and others

• $175M for Dallas-based Island with funding from Insight Partners, Sequoia Capital, Cyberstarts, and others.

• $150M for San Francisco-based Harness with funding from Norwest Venture Partners, Menlo Ventures, IVP, and others.

Top-10 Market Makers

The 10 companies whose valuations dictate what everyone else thinks they're worth

Our Take

The big story: xAI and SpaceX merged into a $1.25 trillion company, combining AI with space infrastructure. Anthropic raised $30 billion at $380 billion (second-largest venture deal ever), while Waymo pulled in $16 billion at $126 billion for global expansion.

Valuations are scattered. OpenAI trades at $723.12 on secondaries, implying a 61% premium over its $500 billion tender offer. That’s speculation. Compare Databricks ($134 billion, 65%+ revenue growth) and Stripe ($159 billion, profitable, $1.9 trillion payment volume) which are both grounded in real performance. ByteDance ranges $370–491 billion depending on the deal. Tether’s pulled-back fundraising and Ant Group’s slow climb to $103 billion prove regulatory risk still moves markets.

How to Play It

OpenAI's secondary market is the easiest entry. For growth with less speculation, Databricks and Stripe offer strong fundamentals riding AI demand. Waymo's worth a look post-raise. ByteDance depends heavily on entry price; Tether and Ant could be bargains if you're comfortable with regulatory uncertainty.

Top 5 M&A Deals

Here’s five of the M&A deals from the past month that we think matter:

• xAI — acquired by SpaceX for an undisclosed amount

• Cimulate — acquired by Salesforce for an undisclosed amount

• Depop — acquired by eBay for $1.2 billion

• Koi — acquired by Palo Alto Networks for $400 million

• Silicon Labs — acquired by Texas Instruments for $7.5 billion

Top Venture-Backed IPOs

Just 9 IPOs have priced in the U.S. so far in 2026, down 47% from the same period last year, with total proceeds of $2.6 billion.

February reinforces what January suggested: the window is open, but only at the very top. If even one of the mega-names files in H1, it reshapes the whole market mood

• Rank One Computing (ROC) | Denver-based facial recognition and Vision AI company. Raised $24M in IPO.

  20/20 BioLabs (AIDX) | Maryland-based AI-powered diagnostics platform. Raised $40M in IPO.

This Month at R136 Ventures

Ventures from the Valley Episode #24: The Scientist Turned VC Funding the Cure for Aging with Omri Amirav-Drory

What if aging isn't a fact of life but the last problem worth solving? In this episode of Ventures from the Valley, NFX General Partner Omri Amirav-Drory makes the case that longevity isn't just a scientific frontier - it's a $38 trillion opportunity to end the root cause behind 9 out of 10 deadly diseases.

Listen here

That’s all for this month. See you in March!

This email does not constitute an offer to sell, or a solicitation of an offer to buy or sell, any security, investment, or other financial product, and is intended for informational purposes only. Any such offer or solicitation will be made only by means of the Confidential Private Offering Memorandum relating to the particular fund.

Building got cheaper overnight but a lot of funds still operate like it didn’t.

Venture capital has always been an information game. You find the signal early, earn the founder’s trust, validate faster than the market, then help the company scale into an outcome.

AI doesn’t change the existence of that game, but the playing field has shifted so much that many funds (ours included), are rethinking how they operate at a fundamental level.

The reason is straightforward: when building gets cheaper and faster, more gets built. But when more gets built, there’s more noise. And when there’s more noise, the old playbook of warm intros, partner gut feel, and a few reference calls stops being enough.

I think about this through five pillars. Not because I love frameworks, but because these are the five jobs a fund actually does, and each one is under real pressure right now.

TL;DR

• VC rests on five pillars: Sourcing, Winning, Validation, Management, Exit. All five are shifting.

• AI massively increases the volume of startups and the speed of iteration, which means separating signal from noise is harder than it’s ever been.

• Winning favors either massive generalists or sharply specialized funds. The middle is an uncomfortable place to sit.

• Validation is deeper and faster, but also more necessary, because polished demos are now cheap to produce.

• Management shifts toward helping AI-native teams stay lean, ship safely, and pivot without losing coherence.

• Exit discipline matters more than ever when valuations can inflate well ahead of real adoption.

The Five Pillars of Venture Capital

Every venture fund, whether it says so or not, runs on five jobs:

1. Sourcing. Finding promising companies early.

2. Winning. Convincing the best founders to take your capital over someone else’s.

3. Validation. Deciding whether you truly want to invest once you’ve found something interesting.

4. Management. Supporting the company after the wire hits.

5. Exit. Reaching a liquidity outcome, whether that’s an acquisition, an IPO, or another path.

None of these are new. What’s new is the pressure AI puts on every single one of them, and how quickly that pressure is compounding.

What AI Changes, and What It Doesn’t

AI changes the cost of building, the speed of iteration, and the volume of data you can analyze, all at once. That combination expands the opportunity set considerably. It also expands the noise by the same factor, which is the part most people underestimate.

What doesn’t change is the need for judgment. Conviction. Trust. A clear view of where a market is heading, not where it sits today. I’ve watched tools improve every year for decades. But judgment has never once automated itself, and I don’t expect that to change anytime soon.

Pillar 1: Sourcing

Sourcing used to be a human-network game. Warm introductions, founder communities, a handful of visible accelerators, etc. Data existed, but analyzing it at scale wasn’t practical, so most funds relied on relationships and reputation to see deals early.

That world isn’t gone but it’s no longer enough. The startup universe is larger now. Teams are smaller but prototypes arrive faster.

I’ve seen companies reach meaningful revenue with three people, which is genuinely exciting, while also making the sourcing problem harder, not easier.

When anyone can build quickly, the question is no longer “who’s building?” It becomes “who’s building something defensible?” and that’s a fundamentally different kind of search.

Strong sourcing now requires model-assisted research across messy, fast-moving data: product launches, hiring signals, open-source momentum, community traction.

It also requires deep domain understanding, because knowing an industry’s real pain points matters as much as evaluating technical architecture. It requires access to networks and datasets that aren’t easily visible to the public.

And it requires tighter segment focus, because narrower geography, stage, and vertical yields sharper research and better differentiation.

Here’s a practical rule I keep coming back to: the wider your mandate, the more you need scalable systems to cover the ground. The narrower your mandate, the more you can rely on depth. But that depth has to be earned through sustained, real research. There’s no shortcut or substitute.

Pillar 2: Winning

Sourcing finds the opportunity. Winning secures it. These are two very different skills, and confusing them is one of the most common mistakes I see funds make.

If a company is exceptional, you’re almost never the only investor who noticed. By the time a strong founder is raising, they have options, and they’re evaluating you as carefully as you’re evaluating them.

Winning is the competitive layer of venture: brand, speed, terms, syndicate quality, and above all, whether the founder trusts you enough to pick up the phone at 11 PM when something breaks.

The market is polarizing in a way that’s hard to ignore. Large generalists compete with scale, platform support, and global reach, but that game requires enormous capital and infrastructure.

Smaller funds win through specialization, meaning a sharp thesis, fast decisions, and value that’s tangible inside a specific domain. When you already understand a founder’s problem before the first call, they notice. That’s what specialization actually earns you, and it compounds over time as your reputation within that segment grows.

The middle ground, where a fund has a broad mandate, modest capital, and generic value-add, is a difficult place to win from. And it’s getting more difficult every quarter as the poles pull further apart.

Pillar 3: Validation

Once a founder wants your capital, the job flips entirely. Now you have to decide if you really want to invest, and you have to make that decision with enough rigor to protect your fund while moving fast enough to not lose the deal.

This is where AI creates an interesting tension.

On one hand, research and analysis tools have improved dramatically, which means diligence can be deeper and faster than it was even two years ago. On the other hand, that same improvement in tools means a polished product demo isn’t proof of much anymore.

When you can build a compelling prototype in only a weekend, the demo tells you about ambition and taste, but not much about durability.

So the real questions have changed:

• You need to understand data advantage: what proprietary or hard-to-access data makes this product better over time, and harder to replicate?

• You need to understand distribution and workflow fit: where does this product sit in the real operating process of an enterprise, and does it replace an existing step or create a new dependency?

• You need to evaluate team composition honestly: can a lean team ship reliably, handle security, and iterate without breaking things in production?

• And you need to look hard at unit economics beyond the initial excitement, because the real test of a business comes when the novelty fades and the renewal conversation begins.

Diligence speed matters but depth matters more. The fastest “no” is still more valuable than a slow “yes” into the wrong company, and the discipline to walk away from something exciting but fragile is one of the hardest skills in this business.

Pillar 4: Management

After investment, a fund has to deliver on its promises. Things like helping the company hire the right people, building systems that scale, avoiding traps that are obvious in hindsight but invisible in the moment, and potentially raising follow-on capital when the timing is right.

The instinct with AI-native startups, especially for investors who came up in the previous era, is to impose familiar structure. More management layers. More process. More headcount.

I understand the instinct, because structure feels like progress, but with AI-native companies it’s usually wrong. The job is to help build a high-leverage organization that ships safely and scales with discipline, often with far fewer people than legacy playbooks assume.

I’ve watched a twelve-person AI company outperform a sixty-person competitor not because they had better individual talent, but because they designed their organization around leverage rather than headcount.

Where management support matters most is in operating model design for small teams. Things like decision cadence, ownership clarity, lightweight governance that doesn’t slow things down.

From there, it’s about helping teams adopt agentic and AI-assisted workflows that raise output per engineer rather than just adding bodies. Security, reliability, and compliance need to be built into the product early, not bolted on in a panic before a Series B.

And when the timing is right, the fund should be opening doors to the next layer of capital, customers, or distribution. Not too early and not too late.

The best portfolio companies don’t need hand-holding. They need a partner who’s seen the terrain before and can point out where the crevasses are, so they can move fast without falling in.

Pillar 5: Exit

Exit strategy starts at entry price. That was true twenty years ago and it’s still true now, but it’s remarkable how often this gets forgotten when markets heat up.

AI markets are particularly prone to valuation inflation, often well ahead of real, durable adoption. The cycle is familiar if you’ve been through a few of them.

A new technology captures attention, capital floods in, multiples expand on narrative rather than revenue quality, everyone finds a reason why this time the fundamentals don’t apply. And then, eventually, gravity reasserts itself.

For funds, the temptation is obvious: ride the wave, justify the price, assume the next round will validate the entry. A more robust approach, and the one I believe in, is to invest with a longer horizon in segments where AI’s impact is underappreciated and therefore underpriced.

These are the segments where strategic buyers will eventually pay for real capability rather than narrative, because they need what the company actually built.

Exit resilience comes from pricing discipline at the moment of entry, because overpaying is the one mistake you can’t iterate your way out of. It comes from grounding your thesis in industry demand rather than model novelty.

It comes from helping companies build enterprise-grade trust through reliability, safety, and compliance, which expands the universe of potential buyers when the time comes. And it comes from preparing for multiple paths, whether that’s strategic M&A, growth rounds, or public markets.

Optionality isn’t indecision. It’s architecture, and building it early gives you choices when you need them most.

The Constant

Everything I’ve described in this piece, the tools, the workflows, the speed of diligence, all of that will keep changing. A year from now, the specific tactics will look different again.

But the core of venture capital hasn’t moved, and I don’t think it will. The best investors I know, across every era, share the same trait: they can see where an industry is heading before it’s obvious and they have the patience to act on that conviction when the market is loud and the signal is faint. No tool replicates that and no model shortcuts it.

So if you take one thing from this piece, let it be this: sharpen your tools, but sharpen your judgment faster. The tools are available to everyone. Your judgment isn’t.