During his tenure as chair from 1987 to 2006, Greenspan was renowned for his cryptic utterances on the economy, leaving rate-watchers befuddled as to whether they presaged a likely cut or hike. His wife, veteran correspondent , famously that their marriage took time because “he claims he proposed three times before I was able to understand. He was so oblique. It was like his testimony.”

In spite of his long history of obfuscation, however, Greenspan is best known for a fairly unambiguous two-word phrase: “irrational exuberance.” He coined it in a 1996 to the  , a conservative-leaning think tank, titled “The Challenge of Central Banking in a Democratic Society.”

One of the speech’s core points was the notion that pricing logic in an industrial economy dominated by durable goods and materials is far simpler than for a modern economy increasingly dominated by software and services.

“What is the price of a unit of software or a legal opinion? How does one evaluate the price change of a cataract operation over a 10-year period when the nature of the procedure and its impact on the patient changes so radically?” he mused, before turning to that most famous insight.

That insight, if I am translating Greenspan-speak correctly, was linked to the question of how one can establish long-term confidence in valuations of assets tied to fast-changing technologies and business models, like software, where prior notions of unit economics no longer applied.

“How do we know when irrational exuberance has unduly escalated asset values, which then become subject to unexpected and prolonged contractions,” he wondered. It’s a conjecture that 30 years later still has no obvious answer.

Notably, Greenspan’s speech actually predated the most heated periods of the dot-com boom, bubble and implosion, which began in the late 1990s and culminated with the hitting its cyclical peak in early 2000. During and shortly after that period, money-losing e-commerce companies like online grocer and pet supply retailer famously went public at then sky-high valuations before abruptly shuttering. Internet infrastructure providers fared even worse, exemplified by networking equipment maker going from Canada’s most valuable company to penny stock in a couple years.

But while losers lost big, winners eventually eclipsed them. Dot-com-era megastars and , for instance, are now worth nearly $8 trillion combined.

That brings us to one of Greenspan’s other well-known analogies: the lottery ticket.

In Congressional testimony in early 1999, pressed for his thoughts on then fast-rising share prices of hot internet companies, the Fed chair the stock-buying frenzy to playing the lottery. He observed that people have long been willing to pay more for a lottery ticket than their chances of winning would justify, simply because they are drawn to the remote chance of a huge win.

”And undoubtedly some of these small companies, which have stock prices going through the roof, will succeed and they very well may justify even higher prices,” he said. ”The vast majority are almost sure to fail. That’s the way the markets work in this regard.”

Fast-forward to today, and one is easily drawn to apply Greenspan’s analogy to the current AI mania. Once again, we’re seeing unprecedented valuations attached to money-losing companies, many in still relatively nascent stages of development.

In other ways, however, this time it’s not a dot-com lottery ticket redo. For one thing, the companies in which a retail investor might be buying said ticket are by no means small. , at its current market cap, is the sixth-most valuable U.S. public company. It’s priced like a winner, not a wanna-be.

Same holds true for recent valuations for and , both of which have confidentially filed for public offerings likely to debut in coming months. Anthropic hit a $965 billion post-money valuation, while OpenAI’s was recently around $852 billion.

One wonders what Greenspan would say about these stratospheric asset price levels. I’d suspect there are better than lottery-ticket odds that it would be something cryptic.

Related 91Ÿ«Æ· query:

Related reading:

• The AI Startup Funding Boom Is Not A Global Phenomenon
• The SpaceX IPO Filing Looks Nothing Like Those Of The Elite Group Of Tech Giants It’s Hoping To Join

Photo: Dr. Alan Greenspan, former Chairman of the Board of Governors of the Federal Reserve, speaks at the Per Jacobsson Foundation Lecture, October 21, 2007, in Washington, DC. (Photo by International Monetary Fund Photograph/Stephen Jaffe used under the .)

Illustration:

It used to be that if you invested in SaaS, you slept well at night. Returns were predictable because the business model was subscription-based and incredibly scalable: build a horizontal cloud-based platform to target as wide a market as possible, charge per seat and grow by expanding the user base.

1, and their peers returned billions to investors on that model. But now, due to AI, where AI agents are replacing humans as the user (through what the industry calls “headless” models) and upending the per-seat model, the SaaS market has lost its predictability. January’s $300 billion single-session wipeout is a leading indicator that the old SaaS model has passed its peak.

Investors are retrenching and trying to predict what’s next as the three frontier AI companies vault into the public markets at multitrillion-dollar valuations. We would argue that these infrastructure platforms enable the next wave of software innovation: AI-native software that automates and enables the $2 trillion white-collar services market.

Generic, horizontal SaaS, as we know it, is a declining legacy model (like on-premise software before it), but investors still have reason to be optimistic about the software market. That’s because AI-native software is going after a much larger opportunity than SaaS ever claimed and the productivity gains and value creation opportunities are unprecedented. The target markets are vertical industry focused and highly specialized, priced differently and built on proprietary data moats that didn’t exist five years ago.

Death of per-seat pricing

SaaS has always been priced on a per-seat basis. That model evaporates the moment AI agents generate most of the usage. A company that once needed 100 CRM licenses for its sales operations team may soon need just 50.

Technology companies facing that reality have to choose a new path forward beyond connecting people’s workflow: perform and charge for the actual work done (usage) or based on outcomes (ROI). A legal AI platform charges per contract drafted, doing the work of a lawyer. Here the software charges for some fraction of the labor it replaces. A spend management AI-native software application can take a percentage of overages found or a chargeback software application could take a fee on the value of the chargebacks it successfully recovers.

The next era of AI-native software runs on automation and performing knowledge-worker actions, not connecting workers or workflows. These solutions reach beyond IT budgets to much larger labor budgets. The companies that adapt will build faster, deliver more value and command a premium for it.

Horizontal is a liability

Generic horizontal SaaS is the most vulnerable to this changing market. If an entire product is a wrapper around a workflow that an AI agent can now handle autonomously, the value proposition may be greatly reduced. Form builders, project management platforms, SMB-focused CRMs, off-the-shelf social schedulers: these categories are compressing fast and may not recover.

The defensible positions now belong to vertical niche specialists, companies that have built what we call the three “Ds.” Distribution through a recurring and longstanding customer base.

Domain expertise specialized to operate in regulated or complex industries. Proprietary data that drives decision-making and is closely held by customers and inaccessible to frontier models.

When your product is built around the specific workflows, terminology and compliance requirements of one industry, ending a vendor relationship is less about migrating data and more about rebuilding a complex web of experiences, corner cases and historical knowledge. Customers stay not because they’re trapped, but because the cost of retraining, reconfiguring and finding a vendor who understands their world is too high.

The more deeply a company understands the regulatory environment, the operational constraints, and the institutional logic of a specific industry and a specific customer, the harder it becomes to displace.

Legal contract repositories, insurance underwriting criteria, bank loan performance data; once embedded in a model and a workflow, these assets create high switching costs that dwarf anything a generic SaaS contract ever produced. You can export a Salesforce contact list. You cannot export your underwriting logic.

People are part of the product

The model that will define the next decade of B2B software deliberately combines software and services, what practitioners call Human-in-the-Loop, or HITL: pairing agentic intelligence with human judgment at the points in a workflow where it matters most.

Legal, healthcare, cybersecurity, construction, financial services, defense; these verticals are defined by high stakes, regulatory complexity and contextual judgment. Routine and repetitive tasks may be mostly automated, but some portion of decisions will always require human judgement because the cost of errors or omissions is prohibitive.

This solutions-centric customer relationship changes what a software company fundamentally is. When a vendor is embedded in how a client operates, handling onboarding, workflow design, optimization and quality control, it accumulates something pure SaaS rarely achieved: proprietary data, domain expertise and institutional trust. Every client engagement makes the product smarter and each deployment deepens the moat.

This is why the most durable software businesses of the next decade will be built inside verticals, not across them. The companies that understand this will stop treating services as a cost of implementation and start treating them as a compounding asset.

A bigger market than SaaS ever was

Even capturing a small fraction of what projects is a $6 trillion annual productivity opportunity from AI transformation dwarfs the traditional enterprise software market. AI-native vertical platforms no longer just compete for the technology budget, they also compete for the labor budget, the compliance budget and the risk budget. That’s a much bigger pie and a more strategic partnership conversation than any per-seat SaaS vendor ever got to have.

The winners won’t be companies that bolt AI onto existing SaaS products, or that add a services layer as an afterthought. They will be the firms with true subject matter expertise that happen to run on AI-native software. They will collapse the boundary between software and services entirely, building businesses whose value compounds with every customer relationship and every data asset they accumulate.

The AI-native software company is a fundamentally different kind of company than the SaaS era ever produced. And it’s worth considerably more.

is the founder, managing partner and chair of the investment committee at . He launched Lateral with a strategy to allocate first institutional growth capital to independent, owner-operated middle-market businesses underserved by typical buyout firms. Previously, he served as a managing director at , a venture capital and growth equity firm that has invested in more than 300 companies including , , , , and . De Silva also previously co-founded , a marketplace for construction equipment that was sold to for nearly $800 million. He received an MBA from , a master of philosophy from the , and an undergraduate degree from .

Related 91Ÿ«Æ· query:

Illustration:

Globally, robotics startups have so far raised $18.8 billion in 2026, compared to $15 billion in the full year of 2025. The figure also handily surpasses the $14.1 billion raised in the peak venture funding year of 2021, and we still have more than six months of fundraising left.

The impressive rise in funding reflects a marked shift in perception among venture investors about the robotics sector, which was traditionally considered an expensive, asset-heavy hardware gamble. In particular, investors appear to be drawn to startups working on embodied AI, or artificial intelligence with a physical body that interacts with the real world in real time.

Noteworthy recent rounds

The surge in funding is driven by a number of robotics-focused startups raising considerable capital from investors this year. Also, interestingly, two of the five largest raises in 2026 to date have been by Austin-based companies.

Topping the list of largest deals in 2026 so far is Austin-based , a defense tech startup focused on autonomous sea vessels. In March, the 4-year-old company raised $1.75 billion in Series D funding, bringing its total funding to around $2.6 billion. led the round, which set Saronic’s valuation at $9.25 billion — more than double its Series C level in 2025.

Earlier this month, Germany’s , a developer of AI infrastructure for robots to learn, collaborate and operate across real-world environments, said it secured up to $1.4 billion in Series C funding. led that raise.

In January, , a robotics company building an “omni-bodied” brain to operate any robot for any task, announced that it had raised $1.4 billion, tripling its valuation to over $14 billion. That financing came just over seven months after Skild raised at a $4.5 billion valuation. led the startup’s latest round, which included participation from , ’s venture capital arm.

On June 15, Beijing-based , which creates water robots and intelligent unmanned equipment, raised $1 billion in a massive Series A round led by .

And in February, AI-powered robotics company raised $520 million in an extension of its $415 million Series A raise in February 2025, bringing the total round to over $935 million. Existing backers , , and joined new investors, including and manufacturing giant in participating in the extension.

Interestingly, spinout has already raised two rounds in 2026. In March, the Palo Alto, California-based startup closed on a $500 million Series A round, co-led by and . Then in May, it raised another $400 million in a financing led by . The company is developing an AI-enabled industrial robotics platform focused on automating industrial and manufacturing tasks at scale.

Exits

While mergers and acquisitions have been relatively robust with several strategic buyouts, the robotics IPO landscape is a bit quieter, particularly in the U.S.

In China, however, a number of robotics companies have recently gone public. The of , targeting a $3 billion to $7 billion valuation, was considered a milestone for the industry. In March, the company filed for an to list on the , and its IPO was widely expected to spur other startups in the space to pursue their own public-market debuts.

, a startup based in China’s Shandong province that makes lightweight industrial robots, in May listed on the , raising about $86 million. And it did not disappoint. Robotphoenix closed its first full day of trading at HK$53.75 ($6.86 U.S.), up nearly 80%, though shares have dipped to the HK$37 range more recently.

On the M&A front, a number of Big Tech and automotive giants have been aggressively acquiring embodied AI and humanoid talent to anchor their physical automation strategies.

In February, AI-powered supply chain provider acquired , an Austin-based maker of autonomous forklifts and lift trucks.

Skild AI in April that it had picked up the robotics arm of in an effort to deploy its technology to warehouses.

And in May, tech giant entered the humanoid robotics field directly by acquiring San Diego-based . The team was absorbed into Meta’s Superintelligence Labs unit to accelerate training of its foundational physical AI model.

Related 91Ÿ«Æ· queries:

Related reading:

• 5 Startup Sectors Seeing Big Funding Growth
• The Year Of Humanoid Robots
• Embodied AI Fuels Record Robotics Funding In China As IPO Momentum Builds
• Amid Record Robotics Funding, Apptronik Raises $520M Series A Extension To Boost Production Of Humanoid Robot Apollo
• Robotics Startup Skild AI Lands $1.4B, Tripling Valuation To $14B In Just 7 Months

Illustration:

This is a weekly feature that runs down the week’s top 10 announced funding rounds in the U.S. Check out last week’s biggest funding deal roundup here.

This week was not an exceptionally busy one for large funding deals, though we saw sizable rounds in a lively mix of sectors ranging from AI to fintech to quantum computing and cybersecurity. The biggest raise was for AI world-model developer, which secured a $310 million Series B. Venture investors also put money into AI infrastructure and AI models for biotech.

1. , $310M, artificial intelligence: Menlo Park, California-based Odyssey raised $310 million at a $1.45 billion valuation in a Series B round led by . Other investors included ,,,, and . Odyssey develops AI world models that create multimodal simulations of real-world environments. The startup has now raised $337 million in funding to date, .

2. , $140M, fintech: New York-based Chronograph secured a $140 million private equity round led by . The company provides portfolio monitoring, reporting and diligence software for private capital investors, an increasingly important market as private assets continue to grow. The new raise, which it describes as growth capital, brings its total funding to date to $160 million, according to .

3. (tied) , $100M, AI infrastructure: Boulder, Colorado-based Hydra Host raised a massive $100 million Series A led by . A of other investors joined, including ,, , and . The company operates a bare-metal GPU platform that connects customers to distributed AI computing infrastructure. With the latest investment, it has raised just under $119 million to date.

3. (tied) , $100M, cybersecurity: Startups that promise to protect companies in the AI era are also raising massive sums right out of the gate. This week, Santa Clara, California-based Ent.AI emerged from stealth and said it has raised $100 million in seed funding led by. Other investors included,, 1,, and. The company, founded by former executives and members of the Security Copilot team, offers an AI-powered workspace security platform that it says can analyze user and AI-agent behavior in real time to proactively prevent cyber threats.

3. (tied) , $100M, cybersecurity, defense: Arlington, Virginia-based Twenty Technologies secured a $100 million Series B at a $1 billion valuation. The round was led by, with participation from, and. The company develops AI-enabled cyber warfare systems for the U.S. military and intelligence community, helping automate and accelerate offensive cyber operations at scale. Founded by former cyber operators and defense technologists, Twenty Technologies has now raised $138 million to date,. It’s part of a growing wave of venture-backed startups building software for military and national security purposes.

3. (tied) , $100M, quantum computing: Berkeley, California-based Atom Computing raised a $100 million Series C led by that brings its total private investment to date to just over $191 million, . and also backed its latest round. Along with the venture money, Atom also received a $100 million Letter of Intent from the under the CHIPS and Science Act that gives the startup additional public backing in exchange for a minority government stake. The company develops neutral-atom quantum computers, one of several competing architectures seeking to commercialize quantum computing. It is one of several quantum startups to receive sizable funding deals this year, following a record-breaking venture investment year for the sector in 2025.

7. , $65M, biotechnology: Watertown, Massachusetts-based Triveni Bio raised a $65 million Series C co-led by and. Additional participation came from. The company develops antibody-based therapeutics for immunological and inflammatory diseases. It has now raised $272 million total from investors, .

8. (tied) , $52M, semiconductor infrastructure: Menlo Park, California-based AttoTude secured a $52 million Series C led by. Other investors included ,,,, 2, and. The startup develops high-speed interconnect technology for AI and hyperscale data centers and has raised $142 million to date, according to . It comes amid robust funding for semiconductor startups this year.

8. (tied) , $52M, digital media: Beverly Hills, California-based Richard Roths Media raised a $52 million venture round led by . The company says it delivered AI-driven marketing and advertising services for “high trust” industries such as banking, law and healthcare. The investment appears to be its first outside capital, per 91Ÿ«Æ·.

10. (tied) , $50M, artificial intelligence: San Francisco-based Bland AI raised a $50 million Series C led by . The of other investors includes , , founder , and others. The company develops AI-powered voice agents that automate inbound and outbound phone conversations for enterprises, a category that has seen growing adoption as businesses look to replace traditional call-center workflows. It has raised $106 million to date, according to .

10. (tied) , $50M, fintech: Brooklyn-based Interchecks secured a $50 million Series C led by,, and. The company operates a payments platform that allows businesses to manage deposits and payouts through a single API, reflecting continued investor interest in infrastructure that simplifies financial operations. It has now raised just under $79 million to date.

10. (tied) , $50M, artificial intelligence, biotechnology: Menlo Park, California-based Radical Numerics emerged from stealth and said it has raised a $50 million seed round led by, with participation from , and . The startup is developing AI models designed to simulate and predict biological systems, with the goal of accelerating drug discovery and advancing precision medicine.

Large non-US deals:

• The largest startup deal outside of the U.S. this week was very large indeed, and also very unusual. , the Chinese AI chatbot startup that briefly roiled public AI-related stocks in early 2025, reportedly took its first outside financing, worth roughly $7.4 billion. The Series A deal, however, comes with a lot of atypical caveats, notably that investors in the deal didn’t actually receive a stake in DeepSeek, but rather in an LLC controlled by founder , per . Those investors also reportedly face a five-year lockup and receive no voting rights.

Methodology

We tracked the largest announced rounds in the 91Ÿ«Æ· database that were raised by U.S.-based companies for the period of June 13-18. Although most announced rounds are represented in the database, there could be a small time lag as some rounds are reported late in the week.

Illustration:

His investment mandate primarily focuses on early-stage technology companies from seed to Series B that align with or impact the global telecommunications, network infrastructure and enterprise software sectors.

Under his leadership, AT&T Ventures targets investments in software, hardware and infrastructure sectors where AT&T’s network scale and internal engineering resources provide a distinct commercial or technical diligence advantage. Portfolio companies include enterprise and deep-tech firms such as , , , , and .

Prior to his current 12-year stint directing AT&T Ventures, Taneja spent more than two decades working across corporate development, venture lending and investment banking. He previously managed M&A and strategic investment activities for during ownership.

Taneja also served as a director at , where he focused on growth-capital debt and equity investments in mid- to late-stage technology businesses, as well as holding corporate finance and investment banking roles at and .

In an email interview with 91Ÿ«Æ· News, Taneja shares why he believes that while AI has drastically lowered the barrier to building software, it has also shifted the definition of seed-stage technical risk.

The new dynamics, in his view, gives AT&T Ventures an opportunity to differentiate itself by offering immediate, real-world technical validation and network integration rather than just capital.

The interview has been edited for brevity and clarity.

91Ÿ«Æ· News: If startups are building fully functioning apps by the seed round using AI, what does that mean for the traditional definition of technical risk? Is tech risk dead at seed, or has it just evolved into something else?

Vikram Taneja: The old definition of technical risk was “can they build it?” Although not entirely absent at the seed stage, I’d say it is becoming less relevant given the dramatically lower barrier to building software with AI tools.

But what replaced it is actually harder to answer: “Is the tech defensible?” Not just “does it work?” but “does it compound?”

Data moats, proprietary training sets, network effects built into the architecture — that’s the new measure of durability.

In prior cycles, technical complexity alone created some natural protection. As a result, the technical risk conversation has shifted to focus on how a company defends itself over the next three to four years, especially as frontier labs move down the stack into application layers and start targeting entire verticals.

Similarly, the distribution question shows up much earlier. “How can you get this to market?” is increasingly asked at the seed stage rather than later in the cycle.

We’re also seeing increased competition for investors to secure larger stakes at seed that they would have previously pursued at the A round. This is driving investors to be more thorough at the seed stage, and founders have to be prepared to meet higher expectations across the board.

When anyone can use AI tools to spin up a working app in a weekend, product execution happens fast, but moats can be incredibly shallow. At the seed stage, how are you separating a truly defensible platform from a beautifully executed wrapper?

Taneja: In early 2025, we saw a wave of AI wrapper companies built on top of frontier models like ‘s GPT, ’s Claude or LLaMA, and a lot of capital flowed into them. What’s changed is that frontier LLMs have now clearly started to take more of a platform approach — moving into the application layers and beginning to pick off the low-hanging fruit.

This is why defensibility becomes critical in AI investing. No platforms are totally defensible, but on some level, you have to ask that question now at the seed stage.

We’re looking for platforms using proprietary data that can’t be replicated by AI, companies that have embedded deep domain expertise — areas where general-purpose AI still lacks industry context — into their workflows, or highly specialized ecosystems or niche markets that provide another layer of insulation in categories that are too targeted for frontier labs to pursue directly.

Are you seeing a change in the actual headcount or makeup of seed teams? If AI handles the heavy lifting of the initial code, are these founders spending their seed capital on engineers, or are they shifting resources immediately to distribution and go-to-market?

Taneja: There is still an engineering focus in the early stage, as there should be, but we are increasingly seeing product, sales, or partnership roles becoming sought after earlier than in the past. And the reason is, as you stated, that it’s easier to build a working prototype, or even a production-ready application, so the focus very quickly turns to establishing trials with customers or exploring distribution paths to dial in the product features.

For strategic investors like AT&T Ventures, where we often do proof-of-concepts with potential portfolio companies, this is very exciting. We get a chance to work with companies earlier in their formation, can get real technical validation much earlier than otherwise, and can similarly try to find a path to collaborate more quickly.

AT&T Ventures has traditionally played heavily in the Seed to Series B space. If institutional VCs are rushing to seed to grab larger stakes because the tech is mature, how does that change the competitive landscape for CVCs? Are you finding yourself competing directly with traditional multistage funds earlier than before?

Taneja: The makeup of seed rounds has definitely changed. Multi-stage funds used to show up at Series A or B when there was enough traction to underwrite. Now they’re at seed because, as we discussed, the companies are mature enough, and they are trying to find winners earlier in the cycle. So yes, we’re in the same rooms as before.

But I’d push back on the idea that we’re competing directly.

A Tier 1 financial VC’s seed check and an AT&T Ventures seed check are different instruments. They are offering capital, brand, guidance and pattern recognition from backing hundreds of companies.

We’re offering something a financial VC structurally does not: our network teams working with your product in a production environment, oftentimes before we even write the check, for example. That’s free diligence running in both directions. We’re validating the company, but it’s also receiving a real-world signal from one of the world’s largest network operators.

For a seed-stage company that’s already solved the building problem and now needs distribution, that’s tangible value and complementary to what financial VC firms are providing. So that competitive pressure has actually sharpened our value proposition. It forces us to bring more than just capital to the table.

Historically, corporate partners want to see enterprise readiness, security compliance and scalability — things a seed startup rarely has. If a seed startup has a fully functioning product but is still a two-person team, can an enterprise like AT&T actually run a pilot with them, or does the corporate integration timeline become a bottleneck?

Taneja: It starts with strategic rationale. That has always been the entry point for us at AT&T Ventures, and that hasn’t changed. If that is in place, then it doesn’t always require full enterprise readiness to start a pilot. It can be a structured trial or a highly targeted engagement, depending on the company’s stage.

We have a number of ongoing proof of concepts with portfolio companies across areas such as AI-RAN, connected infrastructure and computer vision.

The key is clarity upfront — clarity on what the objective of the engagement is and how we measure success. Once that is clear, even early-stage companies can be integrated into a learning or testing environment without unnecessary delay. The goal is to make the AT&T relationship feel like an accelerant to further adoption.

If seed is the new Series A in terms of product maturity, are you seeing Series A pricing bleed into the seed round? How are you disciplined about valuations when the product looks like a Series A, but the company infrastructure is still very early?

Taneja: Seed pricing indeed looks different than maybe four or five years ago. We’re routinely seeing seed deals priced in the low- to mid-single-digit-million range at about $20 million to $25 million post-money. This is pretty much where Series A deals were a few years ago. But it’s not necessarily unjustified — the makeup and traction of seed-stage companies are much further along than predecessor vintages as we’ve discussed.

We stay disciplined by being explicit about what we’re actually underwriting. We’re not just underwriting the financial return on this round — we’re underwriting the strategic value of the relationship over a five- to 10-year horizon.

Does this company make AT&T’s network more intelligent? Does it open up a new customer segment? Does it validate a thesis we’re building around? Are there commercial opportunities beyond our initial thesis? When you frame it that way, it gives us a longer horizon to work with and provides multiple levers to pull.

And honestly, that’s where our engineering and product teams play a key role. They help us decipher whether the product that looks like a Series A is actually built like one, or whether it’s a great demo sitting on a foundation that hasn’t been stress-tested. That technical read bolsters our conviction when making investments.

A functional AI app at the seed stage still requires massive infrastructure. When you evaluate these early-stage companies, how much does their underlying architecture and how they handle data processing or edge computing factor into your decision?

Taneja: Architecture is a key part of our diligence process. The way we think about it really depends on the ultimate use case. Is it for internal use — i.e., a tool that AT&T will be working with in our environments — or is it something we’d be distributing or incorporating into some form of product offering?

If the former, all aspects of the architecture will be reviewed, and this is most likely to occur throughout trials and proof of concepts as we develop a technical understanding of the application or product. If it’s the latter, then we’re likely most interested in understanding how this product architecture scales over time and what it means from a cost, latency and infrastructure perspective. We love to see companies embracing edge-related technologies, but that doesn’t preclude us from working on applications that use traditional data processing methods.

You’ve spoken before about your interest in “physical AI” and robotics (like Apptronik). The software lifecycle is easily compressed by generative AI, but hardware and physical deployment take time. Does this “seed is the new Series A” trend apply to pure-play software strictly, or are you seeing AI accelerate physical tech and IoT at the early stage too?

Taneja: Physical AI is a sector we’ve been looking at quite a bit, particularly because inference and decisioning in autonomous systems, robotics and connected devices create a very different type of demand profile on networks.

The software layer is clearly accelerating — things like perception, control systems and decisioning are moving faster because of AI (the rounds show it!). That will ultimately help pave the way for the adoption of physical AI. However, the physical deployment cycle still takes time, so you don’t see quite the same level of time compression there.

What is interesting for us at AT&T is the intersection — how intelligence is moving closer to the edge and how that changes the way networks need to be architected to handle those workloads.

Related 91Ÿ«Æ· queries:

Related reading:

• In Charts: Seed Deals Keep Getting Bigger As Odds Of Reaching Series A Fall Dramatically
• Amid Record Robotics Funding, Apptronik Raises $520M Series A Extension To Boost Production Of Humanoid Robot Apollo

Illustration:

Six months later, a new tech came along that shook the market. While only a few customers left and the financials remained strong, the stock went down, fast.

This is the danger boards must confront. Disruption rarely announces itself during a crisis. It often appears when the business still looks strong.

For boards, AI, and eventually quantum computing, and other technologies, should not be treated as another technology trend. These technologies can reshape pricing, customer expectations, cybersecurity, product development, talent needs and the company’s business model itself.

Under this fast pace, evolving tech world, company boards should consider the following three points.

Measure the cost of inaction, not just the cost of adoption

Most boards ask: “How much will this AI initiative cost?” That’s the easy question.

The harder question is: “What will it cost if we are late?”

If a competitor uses AI to reduce costs, accelerate delivery, improve personalization or launch faster products, the cost of delay may be far greater than the investment required. The company may lose pricing power, customer loyalty and market relevance before the damage fully appears in the financials. Every major technology discussion should include a “cost of inaction” analysis.

What happens if the company is 12, 18 or 24 months behind? Which margins come under pressure? Which customers become vulnerable? What market image will I have that will impact future clients? Which parts of the product become commoditized?

Challenge the business while it still looks successful

Boards often become more aggressive only when performance weakens. By then, options are limited. The real test is whether the board can challenge management when revenue is growing, customers are renewing and the strategy still appears to work. Success may blur your vision as to what can go wrong.

Boards should regularly ask: Which part of our business would be most vulnerable if AI (or the next big tech change) made it cheaper, faster or easier to deliver? Which revenue stream depends on friction? Which product feature could become free? Which customer process could be automated by someone else?

These questions may feel uncomfortable when things are going well. That is precisely when they matter most.

Build the company that would disrupt your current company

Instead of asking only how to defend the current model, boards should ask management to design the competitor they would fear most.

What would that competitor do differently? How would it price? What teams would it build? What technologies would it use? Which costs would it eliminate? Would it bypass traditional distribution channels?

This exercise forces the company to think offensively. It pushes management to consider bold changes before they become urgent.

For AI, the impact is already visible across software, services, analytics, support, marketing and operations. For quantum, the timeline may be longer, but the strategic implications could be significant in cybersecurity, finance, pharma, logistics and materials science.

Boards do not need to chase every trend. But when technology changes how work is done at the core, when it changes cost structures, speed of development, brand reputation and distribution channels, it becomes a board-level issue.

is a strategic adviser to tech companies, investors, CEOs and boards, specializing in strategy, growth and M&A. He is a guest contributor to 91Ÿ«Æ· News and a university lecturer on strategy, finance and entrepreneurship. Learn more at and connect with him on .

Illustration:

I sold my AI research company while I was qualifying as a lawyer in the U.K. I built with researchers from , , and who believed in the mission enough to trust a 21-year-old law student to lead the ship.

, when acquired us, it was the first time in its 170-year history it bought a company pre-revenue. Thomson Reuters acquired us for the science.

Getting there was painful, though. Our published papers put the model among the best in the world at legal reasoning, and we trained it for a fraction of what the large labs were spending. We had been a quieter version of “the story,” developing very capable models using novel algorithms with huge capital efficiency.

None of that counted for much in the rooms I walked into. Investors always asked about the product and the traction. U.K. investors passed, and I ended up raising most of our funding in the United States.

I back founders now, and the things I weigh have stayed consistent. As a founder, I was told again and again that science meant little until it was bolted onto a product. That test was wrong then and I believe it is fatal now.

Backing founders in the foundational layer

In the first quarter of 2026, foundational AI startups raised around $178 billion. The market is realizing that foundational AI is where the long-term value sits, and this is the year we may see the exits and IPOs that prove the bet right.

However, the capital and the conviction have also pooled around a few names that were already incumbent. , and took roughly 97% of it, and every other foundational AI company in the world shared what was left.

For a deep-tech founder starting out now, that might push them toward a tempting but dangerous read of the market: that the race is over, and that the sensible move would be to build on top of one of these giants.

I’m looking for founders who move the other way.

Most application-layer companies, built on a model they do not own, adapt to the pricing and access decided for them by the firms upstream, and compete in categories that the same firm can absorb whenever it chooses.

The more durable place to build is the layer underneath. The cost, speed, reliability, interpretability and safety of AI systems remain unsolved and genuine scientific challenges, and they decide what everything that sits above them can do.

A real advance in training efficiency, model architecture and inference cost is the work that will still matter in five years, long after most of today’s wrappers have been priced out or absorbed.

Asking the right questions

So the questions I ask AI founders are:

• Is your technical team, scientist-for-scientist, equal to or better than the team at DeepMind?
• Does the problem sit at the level of the model and the system, or is it one more thing stacked on someone else’s?
• Will the “product” get harder to live without over the next five years, or is it looking to reach for early revenue like every other startup?

Some of the companies that ended up mattering most in the AI era are those that survived this line of thought. DeepMind and OpenAI began as research efforts with no obvious product, and both would have looked uncomfortable to a conventional early-stage software investor. Their importance is obvious in hindsight, but the foundational problem-solvers tend to look unfundable right up until it looks inevitable.

Do not build to look fundable this quarter. Build something that the whole stack will depend on in the future. Hire the best team you can find to do it, and solve the hard, foundational problem while it is still unfashionable.

The deep-tech capital market is slow, and it will keep chasing familiar names for a while yet. The work still comes first, and the founders who dare to do it early are the ones the market eventually has to come and find.

The future lies in deep tech, not in the surface wrappers that pass for most products nowadays.

was the founder and CEO of , which was acquired by in 2024. He is an angel investor and senior director at Thomson Reuters Labs.

Illustration:

The acquisition represents an enormous return on investment for Cursor’s backers. Since its founding just four years ago, parent company raised $3.4 billion from investors including , repeat backer , and and was most recently valued at roughly $30 billion in November, per 91Ÿ«Æ·.

The acquisition gives SpaceX, which raised $75 billion in its IPO last week, a foothold into the enterprise software development market, where AI-assisted coding has taken off and led large companies to significantly pare back their reliance on human engineers. Cursor said in November last year that it had crossed $1 billion in annualized revenue.

Hawthorne, California-based SpaceX has in recent years expanded beyond space exploration to become something of an umbrella company for CEO ’s numerous other interests and ambitions, as the company acquired the social media platform (formerly Twitter) and the AI company . SpaceX shares jumped around 16% on Tuesday following the Cursor announcement.

This year has proven robust for M&A activity involving venture-backed startups, 91Ÿ«Æ· data shows. Through June 16, at least 1,177 such deals altogether valued at $182.7 billion have been announced. That compares with 1,132 deals valued at $106.7 billion in the same period last year.

Related 91Ÿ«Æ· queries:

Related reading:

• SpaceX Shares Close Up 19% After Largest IPO Of All Time

Illustration:

“Silicon Valley has done very well with software, but while software was eating the world, they forgot about silicon,” Barrett told 91Ÿ«Æ· News in an interview.

The firm recently closed a $475 million fund focused on investing in deep-tech startups at seed and Series A. In the decade-plus since its founding, it has built its investment thesis around the idea that breakthroughs in science and engineering — not just software — would create the next generation of valuable companies.

With demand surging for compute, semiconductors and energy, Barrett argues the rest of the industry is now catching up. “We’ve been at it for more than a decade,” he said. “In recent years, as AI is eating software, people are scrambling back to recognize that the energy, semiconductors and infrastructure they operate on all need capital too. We’ve been operating in that regime for a very long time.”

Barrett is originally from Australia and came to Silicon Valley in the 1980s. He’s been coding for 50 years, he said, after developing an early and deep respect for science and engineering as the child of two engineers. His childhood was steeped in punch cards, draftsmen and drawings of control systems and machinery, he said.

“Science lets you follow breadcrumbs from prehistoric plumage to semiconductors. One principle can be applied somewhere orthogonal and create extraordinary value,” Barrett said in a lengthy interview with 91Ÿ«Æ· News.

Barrett went on to found video game developer , joined to build the entertainment browser acquired by , and was subsequently CTO at prior to co-founding Playground Global in 2015.

Playground Global operates a lab in the former Palo Alto Research Building in Palo Alto, California. The location hosts 350 people, including those working at its portfolio companies and others with adjacencies working from the lab.

On a recent visit to the warehouse, I saw various models of robots, materials for aerospace construction, and a model of building powerful lasers to increase the speed of semiconductor manufacturing. The quantum computing startup , a Playground portfolio company, moved in when it had three employees and moved out when it reached 90.

The firm has four general partners. Along with Barrett, they are , the former CEO of and who architected CPUs at Intel that helped computing take off at scale, and who joined the Playground team last year as a general partner specializing in semiconductors; , who has made many investments in biotech, including ; and co-founder , who led the investment in .

What follows are highlights from a wide-ranging interview with Barrett that covered topics including sovereign technology, the need to invest in companies that operate on the physical plane, and why he believes putting data centers in space is stupid.

This interview has been lightly edited for clarity.

Gené Teare: What is the thesis for Playground Global?

Peter Barrett: It is about reducing new results in science and engineering into commercial and societal value. That means operating at the boundary between computation and the physical world. We are very interested in new capabilities of computation driving civilization forward, and that inevitably means operating in the same physical plane that we live in.

We’re seeing in our data a huge amount of funding going into space, semiconductors and robotics. It seems as if the whole venture industry has pivoted to this much broader array of companies. Do you see that as a good thing?

Barrett: We lost a lot when people weren’t investing in things that strike us as important. It is good that there is capital chasing the things we care about and that have real consequence.

You can’t spin up a deep-tech practice overnight. You still need domain expertise. You still need to understand why investing in nuclear reactors is good, and why data centers in space are preposterous.

Silicon Valley hasn’t been very efficient with much of the capital it’s deployed over the past decade or so. But I do think it’s good that people recognize that software may be eating the world, but you can’t eat software. We have to operate in the physical layer.

Do you think Silicon Valley gets more efficient?

Barrett: We need to do the work. You develop the instincts and the platform to deploy capital efficiently into these places.

It’s important that people recognize there’s this unprecedented funnel of technical change. AI is an early indicator of it, but we have technologies like quantum. We know how to produce computation using things beyond transistors and semiconductors.

We’re scratching the surface in terms of AI models. We’re right at the beginning of an explosion and renaissance in materials science driven by things like quantum computing.

Now would be the time — and candidly, I feel the imperative — that anywhere there is science and capital, it needs to be turned into value, especially in liberal democracies, because the despots are doing a pretty good job of it. It’s incumbent on us to stay ahead.

We’re in the DOS age of AI. We’re scratching the surface, both in terms of the models we make and the hardware we run them on.

Now would be the time for people to write checks into things that are sensible and valuable. We spent a lot of time on NFTs. How are we doing with cancer? How are we doing with our most difficult challenges in terms of healing and feeding the world?

There are lots of new degrees of freedom that could take capital and turn it into value.

Do you think deep tech fits the venture thesis, despite the long time horizons and the amount of capital it requires?

Barrett: The long time horizons certainly exist. If you’re building PsiQuantum, we’re building million-qubit quantum machines. That takes billions of dollars and a decadal effort.

The corollary is that we’ve had hardware exits in two years. The timelines for hardware aren’t necessarily that different from software.

Therapeutics naturally take a longer time, because of clinical trials. But we’ve also seen exits there. One of our companies tested half a million drugs in a single animal and created a new corpus of AI input for building models to create therapeutics. That’s not a decadal effort — that’s a handful of years before exit.

We try to craft a portfolio that’s a mix of tactical and strategic. Some of these companies get to hundreds of millions in revenue within a few years. Others, like PsiQuantum or , may take a decade to reach full entitlement. That’s part of portfolio construction.

The biology company you mentioned — what’s its name?

Barrett: . It did the largest pharma deal of its kind last year with . The deal could be worth $2 billion on the back end.

It’s a unique mechanism to create giant AI training sets by using physical systems — using animals and in vivo testing to create that dataset. It affords the ChatGPT and biology moment, where you can have large enough training sets to build big models.

You describe the firm as investing somewhere between improbable and impossible. Are there companies that really fit that thesis when you first met them?

Barrett: When we first met PsiQuantum, they were talking about building a machine which was 10,000x the state of the art. Using then-current technologies, it would have been the size of the Sierra Nevadas.

They required exponential improvements in both hardware and software, and they’ve achieved both. It’s the size of a warehouse, not a laptop.

The work we’re doing in biology, materials, quantum algorithms and superconducting logic — which will replace transistors and semiconductors — all of these things sound like science fiction, but they’re much closer to improbable. In many cases they’re entirely practical before we invest; they just seem improbable to those unfamiliar with the domain.

There are things that are not impossible but are still really dumb — data centers in space, small modular reactors (SMRs), or fusion. The physics may work, but the economics don’t, or the timelines don’t align.

I’m disappointed we haven’t invested in anything that turned out to be more impossible than we thought. None of our portfolio companies failed because the technology didn’t work.

We’ve had capitalization failures. We flew hydrogen planes. We’ve built things that were thought to be virtually impossible that turned out to be straightforward. They may have missed their market or may have been unable to raise the capital to continue.

I want to do something where the technology doesn’t work, and we’ve yet to do one of those.

Is there a company you missed out on where it looked impossible and you wish you’d invested?

Barrett: I wish I hadn’t taken ‘s word for it when was a non-profit.

We haven’t missed many. As the roadmap developed, we wish we had been earlier in a couple of categories that are really interesting. But overall, we haven’t missed too many.

In which sectors or companies have you invested where the time horizons have shortened due to AI?

Barrett: Adding Pat Gelsinger to the team reflects an interest in scaling semiconductors along various dimensions, including energy efficiency and how power is delivered.

We do everything from nuclear reactors all the way through to transmission, energy conversion outside the data center, inside the data center, under the chip, what kinds of chips you’re running, what models run on top of those chips, what architectures those chips are made from, and what materials those chips are made from.

At every layer of the infrastructure — optical interconnects, memory systems — we have a best-in-class company at every point. We built the first AI accelerator a decade ago, and we’ve broadened that to encompass the entire ecosystem, from the creation of electrons to how they expend themselves doing useful software work.

There are bubbly aspects of the current AI moment, but the bubble is being modulated to some degree by the unavailability of energy.

We’re in the DOS age of AI. LLMs are embarrassingly incompetent compared to what comes next, but we believe in the durability and growth of AI, and are making investments in model architectures and the ways AIs are trained. We see demand for compute, energy and infrastructure continuing to grow.

We have technologies that can reduce general-purpose compute workloads by 100x to 1,000x over state of the art. We believe we know how to make the energy and deliver it. We know how to connect these systems.

So quixotic pursuits like putting data centers in space are unnecessary.

Talking privately to hyperscalers and Fortune 50 companies, they all say there is way more demand for AI in its future incarnation than exists today. It’s incumbent on us to figure out how to do it 100x, 1,000x or 10,000x more efficiently, because that demand turns into GDP growth and better solutions to our hardest problems.

What are the companies in energy and semiconductors that you are betting on?

Barrett: One example is the wild superconducting logic company . We can make things that are post-semiconductor and post-transistor, with devices that switch five orders of magnitude more efficiently than transistors.

They operate at cryogenic temperatures, but quantum computers do that, and our extreme ultraviolet lithography system does that. The future of computation is cryogenic. Even after you pay to make it cold, you’re still 100x to 1,000x more energy-efficient on compute.

This technology has been around since last century, but it’s mainly been used for secure signals intelligence and radar applications. We’re generalizing it for compute.

Another example is . People talk about SMRs, which are a physics solution to a financial problem, or fusion, which is still decades away. Alva instead uprates the existing nuclear fleet to get hundreds of megawatts out of each unit by replacing 1970s steam generators with a 2020 steam generator.

We can deliver power in a handful of years. No new fuel, no new regulatory path, and a business model that makes sense for operators. We can put gigawatts onto the grid without moving a fence line of an existing reactor and without upgrades to the electricity grid.

We know how to make AI training wildly more efficient. We know how to train different kinds of AI models that we’ve been unable to train.

The last supercomputer at uses something unlike a CPU or GPU to run existing software. We’ve been running software the same way for 70 years, but there are other ways, with dataflow architectures. We have a company doing that — [].

The degrees of freedom from materials, systems, code and models have never been greater. We’re exploring all of them. But most require rolling your sleeves up in the physical world.

LLMs feel like brute-forcing something — like a drunk looking for keys under the streetlight. We’re pushing more and more into that, and I think that’s a dead end. We know other ways of moving forward.

Are you seeing new model companies, separate from LLMs, that are going to solve things?

Barrett: Our brains are not LLMs. They’re not transformers. Transformers are effective, but they are one of a long line of soon-to-be-extinct models that get replaced by something that works better.

That millionfold gap between our brains and GPUs is an architectural gap. Meat is much worse at computation than hardware can be, so biology shouldn’t be better.

Physics allows a million times a million more efficiency, and we should start chipping away at that.

Intelligence is useful and can be pressed into service against basic things like photosynthesis. Plants were invented by accident of evolution 3 billion years ago. They’re pretty, but not efficient. They shouldn’t be green; they should be black. We know how to make photosynthesis twice as efficient, and probably 5x more efficient.

We’re not stuck with the physical constraints of our technology or of nature. Nature is beautiful, but cobbled together by a process that we can have agency over.

All the materials that operate our civilization are discovered, not designed, because we can’t design things we can’t simulate. Our best computers cannot simulate the quantum nature of nature. That’s about to change.

We’re stumbling around in the dark, relying on serendipity and the occasional magical material. Whereas we can construct any number of materials with magical properties that are currently hidden from us by our inability to simulate the quantum mechanical processes that animate chemistry.

We are right on that threshold of unlocking all of these dimensions. And at the same time, we’re putting money into NFTs, the metaverse and other things that will come and go, without anybody ever caring.

Are you talking about the mix of quantum with biology and model-focused companies?

Barrett: Quantum allows us to directly design materials, directly explore the method of action of drugs, and directly design drugs.

AI has a role to play in biology and understanding structures we can measure. We think there are quantum wet labs where we can measure the performance of small-molecule drugs against models of nature and then verify in nature.

We don’t know how many things that animate our industry actually work. We don’t know how Tylenol works. We don’t know how the Type II superconductors we’re building fusion reactors out of work. We know that if you take iron and nitrogen and arrange them in a certain way, they produce magnets stronger than rare earth magnets, but we don’t know why.

There are mysterious things we’ve stumbled across that hint at an Aladdin’s cave locked behind a wall of computation. That wall is coming down.

Which sectors do you think are going to take a lot longer to come to fruition?

Barrett: Civilization will operate on fusion eventually, but right now the only reactor that works using gravimetric confinement is the sun. I think that’s a long way off.

Data centers in space are stupid. You can’t operate a gigawatt data center in a thermos. We have terrestrial answers to those questions that we should pursue.

I’ve always been a detractor of self-driving cars, which are starting to work. Now we need an economic model that makes them sensible and doesn’t drown our cities. The problem with transportation in cities is not the degree of autonomy. If we cared about traffic deaths, we’d worry about roundabouts.

There’s also nonsense with NFTs and the metaverse which have sopped up enormous amounts of capital. Small amounts of capital using these tools against our most difficult diseases would yield results. Small modular reactors are an unwarranted innovation.

There are lots of things that, at first blush, seem good and valuable, but there are far better solutions that are simpler and more imminent. We need to be practical about where the money goes.

There was a company that just joined the 91Ÿ«Æ·, valued over $1 billion this past month, doing orbital data centers. Are you saying this whole category doesn’t make sense?

Barrett: To his credit, will show you a picture of what a 100-kilowatt data center looks like, and it’s bigger than Starship. A 100-kilowatt is a small rack from that is human-sized.

The arguments are that there are a lot of renewables in space. But there are a lot of renewables on the ground too. North Western Australia has solar and wind that are 70% naturally firm, and on the ground, so you can build things on it.

Put a data center in North Western Australia, which we are doing. We have a renewable site 35x the size of Manhattan.

Energy generation and compute in space is a nonstarter because space is not cold. You’re building things in a thermos and need to get rid of heat. A single human-sized rack is 100 kilowatts, which is about the size of the International Space Station’s radiators and solar panels.

Starship has yet to actually put anything in orbit. It’s made some fireworks, which are pretty, and it’s a beautiful thing. is an amazing company because of Falcon 9 and Starlink. But data centers and power generation in space makes no sense.

We know how to build arbitrary amounts of energy generation on the ground with very safe, very large nuclear reactors. We’ve been doing it for decades.

For all the talent and genius rattling around the Valley, we do spend money on silly things.

Do you think now is the most exciting time to be investing, or have some of those investments already been made and are going to come to fruition?

Barrett: We’ve already made investments in things on a really steep trajectory.

Snowcap will take a decade before we’re building GPUs with that technology, but we’ll have commercial product from them next year. We’re getting better at early, undeniable signals.

PsiQuantum is a long journey, but some things just take that amount of time.

X-Lite seems like a ridiculously long journey, although we’re building the prototype facility now, and it received the first money from the new CHIPS Act.

Some hardware companies making silicon or systems are getting significant revenue in a handful of years.

There’s a sleeper in Fund I. Its first trick was to make MRI machines 100,000x more sensitive, and they’re shipping those. In the background they’ve also been developing that core physics to build a new quantum computing modality. So we actually have two quantum computing companies in Fund I.

Even though that’s a 10-year-old company, there are about to be two companies, one of which will be a unicorn virtually overnight.

There are wild things bubbling under the surface that people are going to wonder where they came from.

Companies like — the only co-packaged optics on TSMC — we’ve been working on that for a long time. Now people are waking up to silicon photonics and co-packaged optics.

There are also stealth companies that are indistinguishable from magic. Some of those will come out of stealth this summer.

Is there anything we haven’t chatted about that you think is worth noting?

Barrett: It’s a sobering note, but globally there is a need and desire for sovereign capability in tech — in Western Europe, Australia, Canada and elsewhere.

There are extraordinary pools of capital, pension funds and Australia’s superannuation fund. Given the things we can invest in, globally the West needs to do a better job translating that capital into societal and economic value.

The safety and durability of liberal democracies depends on creating wealth and staying ahead.

We see a resurgent desire to do that in Europe and Australia. Around those pools of capital, there’s ambition. We need to drive that ecosystem globally, not just in the U.S.

The pace of innovation in Ukraine, driven by need, is indicative of changes that can be made in parts of the world less friendly to the tenets we hold dear in liberal democracies.

We can’t operate under the assumption that everybody clever lives in Palo Alto or that we can only invest in things we can drive to. We need to deploy capital globally, and we do. We’re going to do more of that.

Do you feel encouraged by the amount of infrastructure build-out that’s going to happen over the next few years? It feels like it will create a boom in all sorts of technologies because the drive for efficiency will become much stronger.

Barrett: LLMs are not the end. We’ll run LLMs on these data centers initially, but we’ll run their descendants and other more useful things on these machines and on quantum machines.

It’s going to be hard to overbuild because computation is incredibly useful. There’s no upper bound. We’re not in a Malthusian zero-sum game for resources.

We know how to make everything more productive. We know how to grow GDP arbitrarily large. But we need food, energy and medicine there, and we need to normalize the distribution of wealth.

There is unbounded abundance we can unlock if we spend capital on the right things. We know how to do much more of that than people suspect.

The fact that sensible people are considering data centers in space indicates they’re not paying attention to the things we already have in hand that can move the needle.

We do need compute in space. We need AIs in space, sensing in space, and Starlink is great. But we need to use technologies that make sense, not try to make skyscrapers out of toothpicks.

91Ÿ«Æ· queries:

Related reading:

• Sector Snapshot: Quantum Computing Startup Investment Slows In 2026 While Public Markets Hold Strong
• Sector Snapshot: Semiconductor Startup Funding Still Running Hot

Illustration:

For decades, the SaaS playbook was clear: predictable revenue streams, very high gross margins, efficient customer acquisition and strong net revenue retention made a startup very attractive to investors. These metrics built unicorns and defined how investors valued SaaS investments.

But today, with the launch of LLMs, and in the shadow of the “SaaSpocalypse,” 30 years of relative SaaS stability has been shattered and the playbook is being rewritten with disappearing ink.

If you’re a SaaS founder — especially one raising capital — this may lead to uncertainty and confusion. You may lose sleep because the whole market trajectory is uncertain. Investors themselves are trying to anticipate how the SaaS business model will change and, ultimately, what your company should be. To add further confusion, the model many VCs are championing (SaaS and services, anyone?) doesn’t look anything like traditional SaaS. So, what should a founder do?

Ignore the SaaS du jour

In a recent , partner argued that the next trillion-dollar company will be a software business disguised as a services firm, one that sells both tools and outcomes.

His logic is straightforward: For every dollar spent on software, six are spent on services. Meanwhile, LLMs are commoditizing many AI-native SaaS products before they even have a chance to scale. In this world, Bek argues, judgment — not software — is the scarce asset and customers will eventually pay for outcomes, not seats.

For founders, advice like this can be seductive to take. If software margins are compressing and AI is eroding moats, why not follow the trend, add services and open new revenue streams?

Founders need to be careful about taking this fashionable advice because it is greatly driven by investor anxiety and not as much by market reality. What VCs are really responding to are two separate concerns: how to reduce the risk that a portfolio company is disrupted by foundation models, and how to adapt to a new SaaS economy where software alone may no longer command the margins, defensibility or growth premiums it once did.

Founders should instead be prepared to answer this practical question: which parts of their business still matter, which parts have changed, and how do they need to adjust in that context. If the offering is not core to the operations of the enterprise, a pivot will likely be necessary.

The market reset is real, and yes it affects your pitch

The growth-at-all-costs mindset is gone. In its place, investors are laser-focused on capital and sales efficiency, gross and net retention, as well as Rule of 40, gross retention, CAC payback and burn multiple.

What this means for your pitch: A strong SaaS founder today must be able to demonstrate a sharp wedge, a clear buyer, strong usage, measurable ROI and a product roadmap that expands from point solution into platform.

The bar has moved from “Can this company grow?” to “Can this company grow efficiently and organically, retain customers through budget scrutiny, and compound value as it scales?”

AI startups can grow at unprecedented rates, but early hypergrowth can be misleading when switching costs are low and retention is unproven. Investors are excited by AI growth, but increasingly skeptical of AI novelty.

A demo is not enough. You need to prove AI creates durable workflow ownership, not temporary experimentation. Remember, if it takes less than a year to create a company using the current tools, without a sufficient moat, it will take even less time to create an even better company to compete with this one in 12 months.

Focus must be on creating a system of intelligence or a vertical operating system for an enterprise. Understanding workflows is critical. Features and functionalities are no longer sufficient.

Your pricing model is going to change

Seat-based pricing is no longer always the right answer. If your AI performs work independently, customers don’t need more seats to get more value. This is pushing the market toward usage-, consumption- and outcome-based models. notes that long-term pricing is shifting toward value-based and outcome pricing, and that continued cost-of-intelligence improvements could eventually help margins expand.

In the old SaaS model, value was tied to access: seats, users, departments. In the AI era, value is tied to outcomes. Software isn’t just helping employees do tasks anymore. It’s beginning to execute them directly: writing code, reviewing contracts, resolving support tickets, analyzing financial data, automating back-office workflows.

Have a big moat

Promising AI categories are attracting 2x to 3x more competitors than in prior years, while large SaaS incumbents are aggressively launching AI products, acquiring startups and hiring AI talent. Investors will ask you directly: what’s your moat? Is this a real defensible position, or a feature that 1, or can ship in a quarter?

AI expands the addressable market for software significantly. Traditional SaaS captured software budgets. AI-enabled SaaS can capture services spend, labor spend and outsourced process spend. Battery frames this as a major expansion from cloud software into services automation and human labor displacement — a much larger opportunity than prior SaaS waves.

Rules for the road

The market is open for exceptional SaaS companies. But the bar is higher, and investors have seen enough AI pitches to be skeptical of the theme. What they want to hear from you: A specific customer pain point with evidence of urgent demand; proof of retention, not just initial adoption; efficiency metrics that hold up under scrutiny; and a clear, concrete explanation of how AI improves your product, your business model and your customer’s ROI.

The founders who get funded in this environment will be domain experts who understand their customer’s workflow deeply, where AI can safely replace, augment or accelerate human work, and disciplined operators who understand the economic tradeoffs: when to use frontier models, when to use smaller specialized models, when to fine-tune, and when to preserve human review.

is managing partner at . He has more than 41 years of C-level global experience in the tech sector as a seasoned investor, entrepreneur, board member and educator focused on helping teams prepare for rapid growth, scaling and liquidation events.

Related reading:

• Why The Future Of SaaS Is Eat Or Be Eaten
• 3 Ways AI Is Transforming Venture Capital Investment

Illustration: