AI Ascendant

First Principles

Many of us have been having the "automation vs augmentation" debate for years. Will the most important effect of AI be the replacement of human labor, leading to greater business efficiency alongside mass job displacement? Or will it be the enhancement of humans with AI-derived superpowers that drive huge leaps in productivity? The answer is probably "both". But there is also a third source of economic value from AI: new discoveries. In fact, this is what got most of the attention in the early days of AI. Researchers were excited about new inventions, scientific breakthroughs, solutions to previously unsolvable problems, and unimaginable new human knowledge. Now we just seem to be talking about reducing the need for paralegals, brokers, and programmers. "The TAM is All of Human Labor" is one depressingly clinical take on the situation. This has gotten so much attention that we may be losing sight of what is most special. Are we once again settling for 140 characters in lieu of flying cars?

What is expected of humans needs to change for us to thrive in the age of AI. Our educational system has become overly focused on the "how," training students to execute well-defined tasks like financial analysis and programming. With machines doing amazingly capable work in an ever-expanding set of tasks, the "what" and the "why" need to become humanity's focus. Now we can focus our human energy on strategy, imagination, creativity, moral reasoning, and emotional intelligence. This is both a luxury and an imperative. A world where we spend more of our time on higher-order, higher-impact endeavors seems like a more fulfilling one. And if we don't learn how to equip ourselves to do so, we may face a grimmer reality.

We've been dismayed to hear that some young people are choosing to abandon knowledge work in favor of blue-collar vocations they believe to be "AI-proof." This may indeed be the right choice if that is where your passion lies, but choosing a career simply to get out of AI's way leaves a lot of joy on the table. Rather than resigning ourselves to diminished prospects, we should seek an explosion of opportunity in which many more people can flourish as scientists, inventors, builders, creators, and leaders. Will there be disruption in the transition? Absolutely, and navigating this will be one of the largest challenges of the next decade. But our youngest should be the most optimistic, as they will be best able to take advantage of the opportunities of the new era.

We at Wing are certainly optimistic. In our role as investors and company builders, we see a slate of incredible opportunities made possible by advanced AI. Legacy businesses already face tremendous pressure to adapt or be left behind. New AI-native businesses are already rising to take their place. This clearly applies to technology companies, but the same forces bring much of global GDP into play.

Some Closer Looks

Let’s unpack some of these areas of opportunity in a little more detail:

The Complete Reimagination of Software

There has been a lot of hand-wringing about the future viability of today's software businesses, especially B2B applications. Much of it is justified though the timeframes for the anticipated atrophy are probably longer than some expect. Today's software industry is under threat from at least two directions: reduced software purchasing as businesses build more of their own tools using AI coding agents, and competition from new AI-native attackers.

The "Return" (and Expansion) of Internal Development

There has always been a large part of the market that develops its own software. Early in the IT revolution, this was almost 100%! Paying for software (beyond what came bundled with the computer system) was a novel concept. Eventually the modern software industry emerged and the balance shifted towards purchased products. Now that AI has made it so much easier for customers to build their own software, specifically tailored to their precise needs, the balance will shift back to internal development. Wing is excited by this renewal of internal (and even personal) software development, in a far broader and more democratized form than anything we've seen before, and we are very interested in new companies like Wing portfolio company Dust that help enable this movement.

AI-Native Software Invasion

Will the pendulum swing back to the nearly 100% internal software development of the early days? There will always be customers with a preference to buy, and there will always be categories where internal development is either too big an effort to justify or faces ongoing disadvantages relative to commercial products. In these categories, there is a recurring pattern. Entrenched pre-AI incumbents are struggling to replatform themselves on a Data+AI foundation, all while not trashing their business models and market caps. Meanwhile new AI-native software businesses are rising up with agentic capabilities and disruptive business models. Increasingly, the form these companies take is AI coworkers: products that can do more of the work than past generations of software and, in the process, attack and capture areas of spend that were previously unaddressable by software. In some categories, this new competitor may be a frontier lab itself. A lot of value is going to shift from old to new vendors, and a lot of net-new value will be created too. Wing is very focused on supporting these new AI-native competitors, particularly those with proprietary, accumulating data advantages. Wing portfolio company Gong, the AI OS for revenue teams, is a great example. From the outset Gong has been built on a highly distinctive foundation of proprietary, enhanced customer communications data. That data foundation is what makes enterprise-grade agentic GTM a reality for Gong's more than 5,000 customers.

The AI-native software business is going to have a very different financial model than what we've become accustomed to. Pricing will of course be different, shifting to consumption-based and outcome-based models. But perhaps the most profound change is that, for the first time, the software industry now has a high-cost consumable input: compute. One of the most attractive things about the old software industry was that delivery of customer value was largely decoupled from the cost of that delivery. No more! Now value is directly tied to compute (or if you prefer, tokens, or energy — all directly related elements of a tightly synchronized supply chain), an expensive variable cost that is a fact of life in many industries but a new concept in software. We will need to develop a new way of understanding and evaluating the financial attractiveness of AI-native software businesses, much as we needed to develop new frameworks for the original independent software vendors (which looked so different than the semiconductor, computer and disk drive companies that preceded them), and then again for the SaaS businesses that succeeded them.

Vertical Agents and Applications Unleashed

An exciting consequence of AI and agents is the ability to address vertical markets that had previously been unattractive to build for. Examples include law and medicine, sectors once considered graveyards for startups. Customers in these markets have historically been slow adopters of information technology and paid very little when they did. But this has changed rapidly in many vertical markets with the introduction of AI-native products. This undoubtedly is driven by the superior customer value being delivered for workflows untouchable with pre-AI technology. But ease of adoption is a factor as well, especially for agentic products with simple voice and chat interfaces. We expect to see the emergence of new leaders serving many of these previously underserved markets. Again, the most obvious plays will be attractive targets for the frontier labs' own agentic offerings (e.g., financial services, health, law) but that doesn't mean there isn't also great opportunity for startups with unique perspectives. Wing has a particular focus on the health vertical, where we are helping to build companies like Rula (mental health).

AI-Native Industries

Is selling software even the right business model anymore? The rise of AI has brought this first-principles question to the fore. In many cases it may make sense to forward-integrate and truly deliver outcomes, a "full-stack" business operating not as a technology vendor but as a competitor. This is not an entirely new concept: examples of tech-enabled services can be found in finance, retail, communications, transportation, and other very large markets. But AI opens up many more such opportunities in a broader array of industries. Benefits include superior value for customers by taking ownership of more of the value proposition, as well as superior value capture for the vendor. This all needs to show up in the financials of course, and once again founders and investors will need to develop a new way of evaluating the performance of these businesses. Hallmarks of success will include operating leverage relative to headcount (indicative that there really is leverage in the AI), margins (indicative that this isn't just a token resale vehicle) and go-to-market efficiency (which may be hard in some markets with high historical stickiness, causing some new entrants to consider roll-up strategies). Wing is intrigued by the potential of this model and is already helping build such companies in industries like specialty lending (such as Drip Capital and Juvo) and health insurance (such as Angle Health).

The Unfathomable Infrastructure Requirements of AI

The founders of Wing have spent decades working on next-generation infrastructure, dating back to our days building early GPUs and high-performance networks. We've worked as operators and investors across all the major infrastructure sectors for most of our careers, and lived through the characteristic boom/bust cycles of prior eras — including the frenetic broadband telecom buildout of the late 1990s and early 2000s. Carrying that scar tissue, it would be easy to look to history and forecast an AI infrastructure bubble. Someday that may be an accurate projection — but not today. We are still in a relatively early phase of the largest infrastructure buildout since electrification.

The Compute Consumable

One reason for this belief is that for the first time, software now relies on a scarce, high-cost consumable: compute, and the directly related elements of the same supply chain, including energy and tokens. Continued operation of our newly intelligent systems requires continued consumption of compute. Increasing usage — or increasing intelligence — requires access to additional compute resources. And those resources need to be refreshed every few years; Moore's Law may have slowed in general-purpose computing, but in AI computing the performance and efficiency gains from each new processor generation remain steep enough that the ROI on upgrading is unusually high relative to prior infrastructure cycles.

Your view about the future of AI infrastructure ultimately boils down to your view about the future consumption of AI. There is the possibility, of course, of a major breakthrough in efficiency that breaks today's iron linkage between compute and capability — and Wing is investing according to that thesis in several ways (our portfolio company Unconventional AI is one example). But even with that possibility, we don't see a slowdown in demand for either AI or its infrastructure in the medium term. The hyperscalers' own capital allocation confirms this: annual global data center investment is racing towards $1 trillion when hyperscaler capex, sovereign programs, and purpose-built AI cloud providers are included, and it is still not enough. Wing portfolio company San Francisco Compute both benefits from and enables this demand, creating a more flexible, lower risk way for customers to obtain the computational resources they need. The largest infrastructure spenders in the world are all sourcing capacity from third-party providers they cannot build fast enough to displace. When that is true of the companies with the largest balance sheets on earth, the demand signal is unambiguous — we need more AI compute.

The public markets have finally caught on to the magnitude of this opportunity, opening their eyes to well-positioned companies across the AI supply chain beyond NVIDIA — which itself relies on more than 200 supply chain partners to deliver its flagship AI systems. Some of the biggest beneficiaries are decades-old businesses whose products are suddenly in very high demand — optics, memory, power delivery, semiconductor packaging — delivering venture-style returns in a matter of months.

A New Infrastructure Architecture for AI

The deeper opportunity is in re-engineering these infrastructure products from the ground up around the specific demands and bottlenecks of AI workloads — and the scope of that redesign is unlike anything prior data center generations required. From miles of cable per rack and complex liquid cooling systems, to massive HBM memory requirements, to photonics and THz radios for high-speed data transfer (such as Wing portfolio company AttoTude's "THz Radio Over Wire" technology), the entire system is being rebuilt. The reference architecture is changing constantly, and Wing is investing in companies that push the boundaries of what's possible — across advanced interconnects, networking, inference-optimized processors (such as Wing portfolio company SiMa's low-power edge AI processor), power delivery, memory disaggregation, AI-optimized storage and infrastructure software. As Jensen has observed, an entirely new industrial system is being forged.

Enabling Enterprises and Developers

After a few years of experimentation, we have now reached a tipping point in AI adoption. Enterprises are pushing AI systems into production with real urgency, and developers are launching and improving commercial AI products at an astounding rate. Their needs create opportunities for products that improve their ability to realize business value in these efforts, for themselves and for their customers.

Agentic Development Life Cycle

Defining, building, deploying, and operating agents will be the prime directive in both enterprises and their technology vendors for years to come. There are many unsolved problems associated with these activities, each of which will hold back realization of value if not addressed. Builders need tools throughout the agentic development cycle. This toolkit includes more than just the famous coding agents. We are already seeing delineation of distinct layers and roles in the AI development stack, covering areas like review, orchestration, evaluation, learning/improvement, tool use and observability. Some of these have early leaders that have staked out strong positions, others remain up for grabs, and all are continuing to morph rapidly.

Data and Knowledge

There is a high degree of consensus amongst the enterprise technology leaders we speak to that leveraging their first party data is a top priority. They rightly see this as the key to achieving better business outcomes and sustainable, compounding advantage through AI. This requires construction of a new knowledge layer that organizes that data and makes it accessible to AI in an intentional, secure manner that takes the guesswork out of agentic operations and constricts wild results that would otherwise emerge. Wing portfolio companies Cohesity, Pinecone, and of course, Snowflake underpin such efforts at thousands of enterprise customers.

Inference

Perhaps the largest and most obvious enablement opportunity lies in inference. The first generation of inference platforms has already risen to significant scale serving language models. As new model architectures emerge supporting other domains (e.g., world models, video, robotics) there will be opportunities to optimize around their requirements. If these are distinct enough, new inference platforms will have room to gain scale of their own.

The Rise of the Open Ecosystem

Open weight models are soaring in importance. They offer the hope of cost-efficiency as well as the prospect of AI built for the requirements of a specific application, use case, enterprise or user. We see a broad range of opportunities for products that allow enterprises and developers to best take advantage of them. Examples are found in the various layers of compound AI systems that leverage a constellation of models to deliver superior results at lower costs. There are also opportunities to automate and optimize the adaptation of open weight models in highly targeted, continuously improving fashion. Sometimes this will be delivered as a product, sometimes as a fairly bespoke, FDE-centric service.

Some of the opportunities associated with developer and enterprise enablement will be fleeting and become subsumed by adjacent tools, layers in the stack and service providers. Others will be durable and support offensive consolidation. Telling the difference can be difficult but we are confident there are important businesses to be built in quite a few of them.

New Terrain for Foundation Models

The breakthrough impact of language models raises an obvious question: where else can foundation models change the game? We believe the answer lies in the domains that power the global economy but have never had models built for them — scientific research, physical systems, biological processes, clinical medicine, and the sensory and operational data that drives industry.

In most of these domains, a useful foundation model does not yet exist. And unlike language, where the application layer is separate from the model, domain-specific AI often requires solving the model to unlock the application. The team that does both captures value at both levels simultaneously — a structural advantage that compounds over time as proprietary data accumulates through deployment.

The Quest for Compounding Advantage

Across every category we invest in, the same question determines whether a company has a chance to lead: does deploying the product generate data that makes the model better in a way competitors cannot replicate? This matters because the frontier never stops moving. General-purpose models will continue to improve, and capabilities that look defensible today will eventually be commoditized. The companies with durable positions are not the ones with the best model at a moment in time — they are the ones whose model improves faster than anyone else's because of what they learn through operation. Clinical imaging systems that process real patient data. Robotics platforms that accumulate fleet telemetry across thousands of deployments. Synthesis pipelines that close the loop between prediction and experiment. In each case, the data generated by doing the work is the moat — and it widens with every customer, every use case, every result. Wing companies Deepgram (voice), Goodfire (interpretability), and Tahoe (biology) are all building such advantage in their respective fields.

The best-positioned companies in this landscape are not racing against the frontier labs. They are using them. A better base model means better distillation into a tighter domain loop, which means faster improvement on the metrics that actually matter in the field. That is a durable position. A company whose advantage depends entirely on a benchmark lead is not. Wing portfolio company Intology is building this kind of durable position in the emerging field of AI science.

Data Demand Scaling and Evolving

Developers and researchers developing foundation models have some specialized and often extreme needs. Most obvious of these is data. In domains where proprietary operational data is the moat, outside data suppliers face natural limits, as the model company captures that advantage internally. But in emerging areas like world models and physical AI, where the data requirements are vast and the collection infrastructure is still being built, there is real opportunity for data companies to play a foundational role. As an example, Wing portfolio company Tahoe's perturbative single-cell dataset is already enabling major advances in virtual cell models.

There are other opportunities to supply the labs with the unique resources they need to continue to push their respective frontiers. Wing portfolio company Bespoke Labs works on the frontier of reinforcement learning for long horizon tasks, supplying data and environments to the leading labs as well as enterprises as they push the envelope of agentic function. As we have already seen in the short but amazing history of language model development, this can sometimes support the creation of new businesses of significant scale.

Physical AI

Of the domain-specific AI opportunities, Physical AI may be the largest — a domain as broad as the physical world itself, where the gap between what models can currently do and what is ultimately possible remains vast.

We think about Physical AI in four layers: components (actuators and sensors), infrastructure and tooling, foundation models, and vertical solutions built with the right embodiment for the problem being tackled. At the component layer, costs have fallen dramatically, but the need for better parts persists, particularly in areas like end effectors and touch sensing, and the supply chain realities of the category make U.S.-made components an opportunity in their own right. At the model layer, there is still genuine debate about what the final architecture will look like and the role of vision-language-action models (VLAs), world-action models (WAMs), and other architectures. We are in the early innings, progressing from an era defined by better pre-training before we get to one defined by post-training and reinforcement learning in real and simulated environments.

Supporting these models will be a new class of deployment companies that take a practical view of putting robots into production today: building the right embodiment for the task they are going after, collecting data on real tasks, often through teleoperation, and driving toward full autonomy over time. These are the companies building the know-how and expertise to deploy this new age of robotics. Underneath all of it, there are a number of infrastructure opportunities. The data bottleneck is real. With no internet to leverage, the field needs egocentric data, tactile capture, better simulation and synthetic data including world models so we can kickstart the flywheel of real-world fleet data as deployments scale. We also see other infrastructure opportunities emerging alongside these both pre- and post deployment such as around policy tuning, evaluations, safety, and observability.

Wing has invested in Physical AI companies such as Yann LeCun's AMI Labs, which is addressing this opportunity, as well as data supplier Human Archive, which seeks to accelerate their work.

AI Security

Security requirements accumulate on top of everything that came before. At the base are the primitives built for human decisions and human-speed attacks, and the disciplines that defend that surface are now being stretched upward to cover agents as well. Wing portfolio companies Cyberhaven, which traces where sensitive data moves, and Theom, which governs what structured data is permitted to reach, are both extending into a world where the actor in question is an agent rather than an employee.

Agentic software arrived on top of that foundation, and it arrived closed. The first agents came from a small number of providers running proprietary models on their own inference infrastructure, with identifiable control points. Open weight models followed and are still spreading, deployable on any infrastructure and without those same control points. We expect most enterprises to run in both regimes for a long time, and the further they go toward the open end, the more the enforcement burden falls to them rather than to a provider.

Underneath both is something genuinely new: software can now act on its own authority. An agent is neither a human nor a service account, and that breaks a core assumption of how security has always worked.

We are looking to invest in two types of security businesses. The first is invention. Some of what now needs securing did not exist before. Wing portfolio company Gray Swan attacks AI systems in order to defend them. An open arena where thousands of researchers compete to break frontier models produces the attack data that trains an automated red-teamer, which enterprises run against their own agents before they ship; what it finds hardens a runtime product that enforces policy in production. Each layer feeds the next, and the frontier labs are already among the users. Incalmo sits beneath all of that, building the high-fidelity cyber ranges in which AI attackers and defenders are trained and measured. Whoever supplies the ground agents train on holds a position underneath everyone building them.

The second is reinvention, where AI changes how the work itself gets produced. Security has always been limited by the number and availability of skilled professionals, so what gets done is a fraction of what is worth doing. When AI turns that labor into software, coverage extends to work no team could afford to staff, and the category gets rebuilt rather than improved.