Think + Insights: A Framework for 2026 AI Investment

Key Trends to Watch in 2026

Enterprise Applications. We expect 2026 to be defined by targeted applications, with AI systems becoming embedded within existing tech stacks to improve efficiency, generate additional revenues, and improve the cost structure.  While we will continue to see some novel applications, we expect the biggest traction to come not from “replacement” models but “improvement” ones.  We consider health care (operations and diagnostics), industrial, and general back-office operations as the most promising sectors.

Physical AI.  This is an area we expect to gain more traction in 2026 as the portability of smaller, yet capable edge models improves and edge compute power increases with new chips. In many ways, physical AI is far more useful and easier to implement than information-based AI, since the scope of activities is better defined and the outcomes are also easily measurable.  At the same time, the benefits are both tangible and very valuable. While most of the physical AI innovations that have gained attention have been on autonomous mobile robots (see Serve Robotics, for example), we expect the next wave of physical AI innovation to be in healthcare (e.g, Shyld.ai), industrial, manufacturing, and even retail use cases, where AI will be embedded in the existing physical system.

E-commerce and Advertising. LLM’s role in e-commerce, and by extension advertising, will be very important to watch and is likely to evolve significantly this year, though we don’t expect a tipping point yet. In 2025, content providers and e-commerce companies started to consider, negotiate with, or file lawsuits against the LLMs. The landscape and the boundaries are still not defined, but may start to take shape in 2026.  While some companies, like Expedia, may be willing to work with LLMs and agents, others, like Amazon, are still guarding their walled garden carefully. Imagine a generalized form of Amazon Rufus that has access to nearly all the e-commerce vendors and their full product data.

Our long view is that advertising in its current online forms is on a path of long-term decline, and new models will emerge over the next decade.

Consumer Apps. We believe consumer applications of AI will be among the most transformative developments.  While enterprise apps will be highly valuable for improved efficiency, consumer apps have the potential to completely change the landscape in both user experience and the composition of key players. However, this will take several years and requires systems, such as full voice-based UI and interoperability, that are not yet ready. For 2026, we are excited about the expected release of Apple’s new Siri and a series of promising voice-based startups that are beginning to shape a new UI.  We will also be looking for new consumer applications in personalized health care, shopping, and content and entertainment.

Agents. Autonomous Agents (bots that independently plan and execute complex goals) failed to meet the expectations in 2025, as the complexity of reliably chaining API actions proved too high for enterprise or consumer trust.   We don’t expect a breakout year in 2026 for general-purpose agents. However, we see promise in workflow agents with well-defined constrained systems, powered by new reasoning models that handle specific, repetitive tasks rather than open-ended goals.

LLMs: The Era of “Trust & Verify.” The most significant shift in 2025 was the move from “Generative” to “Reasoning.” LLMs evolved considerably in 2025, most notably by shifting from ‘generative’ to ‘reasoning and search-based’ models (especially with the launch of the O version of ChatGPT). By the end of 2025, more than 50% of LLM tokens were routed to reasoning models.  This has helped broaden the search applications for consumers and information workers considerably, making the results more reliable and verifiable.  However, the biggest development by far was the launch of Gemini 3.0, which significantly improved the quality of its results and the platform’s usability by expanding the understanding of the prompt’s context and its multi-modal use of the “Deep Think”.  We expect this move toward search-based, verifiable results, through new methods such as Recursive Self-Correction, to gain further traction in 2026.  The next stage is what we call “circular verification”, i.e, to do a full circle check of the links provided by the models as the verifiable source – a task that is still done mostly by human users.

Risks for 2026

1. The Unit Economics Challenge. The cost structure of AI remains the elephant in the room. While token costs have dropped for standard tasks, “reasoning” is still very expensive. Models like OpenAI o3 or Claude 3.5 Opus cost upwards of $60–$80 per million tokens. Even mid-range efficiency models are not yet cheap enough to support ad-supported business models comparable to traditional Search. We believe that for AI applications to achieve software-like margins, inference costs must drop by another factor of 10.

2. The AGI Capital Bubble. The industry’s obsession with “Artificial General Intelligence” (AGI) has driven data center build-outs that may far exceed near-term utility. If AGI is not achieved soon (which we do not expect to happen), the ROI on these massive infrastructure capex projects will turn negative, potentially triggering a credit event that can impact the broader sector.

3. Overestimating the Adoption rate.  While we don’t expect a sudden drop or a notable reduction in the rate of enterprise adoption, we do expect the adoption rate to be slower than most investors are projecting.  Most companies that are using AI are benefiting from it, albeit at low rates. Case studies of massive benefits are few and generally exceptional situations. In fact, several studies indicate that most companies are not yet deriving sufficient benefits from AI adoption, and adoption rates remain relatively low despite the push from CEOs.  We expect this gradual trend to continue, but most of the industry may be aiming at a rapid take-off in adoption in 2026, which we think is unlikely.

4. Social/Economic/Policy factors. This catch-all group encompasses risks that have been around and are likely to remain in 2026, but we don’t consider them especially elevated.  They include a pushback against AI if it is seen as replacing jobs; the power demand of data centers and their impact on the nation’s power supply; and, most seriously, the abuse of AI for new cybercrimes.  It is also possible that a few harmful incidents, beyond the usual hallucinations and infringements, can seriously damage the reputation of AI models and push things back several months or more.  Finally, AI has to reach a point where it can be considered a “trusted” platform.  We note that it took autonomous vehicles more than 10 years, and nearly double what most industry observers believed, to be accepted and trusted.

5. Capital Markets and Business Cycles. The overall U.S. Economy has shown remarkable resilience in the face of major challenges in 2025 and remains on a relatively solid footing.  However, there are two economies in the U.S, and they are clearly diverging.  The upper class and asset-owned families are not only doing well but also continuing to spend, buoyed by record stock market gains.  The lower-income groups with little in assets and net worth are facing high costs and a tough employment market, reflected in a sentiment of serious concern. At the same time, the equity markets are priced for perfection with the highest multiples since the dot-com era.  Any disappointment in earnings or revenue growth from major tech players could trigger a correction.  This, in turn, could gain momentum if the economy is weak, especially if interest rates begin to rise.

Our  (unlikely to attain) Wish List
What would be really exciting to see, but we are unlikely to get it in 2026:

• A truly reliable e-commerce agent (that handles returns/negotiations).
• A proactive health monitoring agent (preventative vs. reactive)
• A functional voice UI for all the apps and transactions
• A new AI-embedded and smart email client
• An interactive and smart news and media aggregator/curator agent
• Peace on Earth (one can dream, right?)