Where AI is headed in 2026
12.30.2025 | By: Ashu Garg
I grade my 2025 predictions and make 8 new ones.
Last year, I published my first set of AI predictions. Now it’s time to see how they held up, and to make a fresh set for 2026.
A special thank you to our portfolio founders – Jonathan (Turing), Ram (Maximor), Ashutosh (founder of Viven and Eightfold), Animesh (PlayerZero), Ishan (Oliv), and Mohit (founder of SciFin, Cohesity, and Nutanix) – whose perspectives shaped much of what follows.
Let’s start with last year’s predictions.
Scoring my 2025 predictions
🟢 Where I got it right:
Reasoning models and compound AI systems would define the year. They did. OpenAI’s o1 and o3 showed that test-time compute – letting models “think longer” – could unlock capabilities that bigger base models couldn’t. Meanwhile, the focus shifted from standalone models to orchestrated systems: multiple models, tools, and verification steps working in concert.
AI would rewrite software economics. Our “services as software“ thesis – AI that delivers outcomes rather than selling seats – is now the default frame for B2B investing. AI startups are targeting the services market, not just the software market.
OpenAI’s dominance would erode. Gemini, Claude, and open-weight models all took market share. Enterprise customers adopted multi-model strategies. OpenAI is still the leader, but no longer the default.
NVIDIA’s near-monopoly would face real challenges. Cerebras gained traction, AMD shipped at scale, and Anthropic made a strategic bet on TPUs. The chip wars are now a multi-front battle.
Robotaxis win public trust. If San Francisco is any indication, they did. Waymo’s white Jaguars are ubiquitous here, and the company plans to 4x its service area next year.
AI-native startups would dethrone the incumbents. Not yet. Google in particular roared back. But the pressure is real – incumbents are playing defense for a reason.
AI interfaces move beyond the chatbox. This is happening, just more slowly than I expected. I’m doubling down for 2026 (more in prediction #8).
RIP 10 blue links. AI Overviews now reach 2B monthly users, and click-outs have fallen sharply. Long term, the real disruption isn’t AI search: it’s agentic commerce, where humans aren’t in the loop at all (see my prediction #5).
❌ Where I was wrong:
Llama would become “AI’s Linux.” Open weights models are thriving, but the ecosystem fragmented rather than consolidating around a single Meta standard. There’s no Linux-style dominance, just a healthy plurality.
8 predictions for AI in 2026
🏭 Prediction 1: Enterprise AI finally hits production
For the past two years, enterprise AI has felt perpetually “just around the corner.” The holdup isn’t model capabilities. It’s that AI can’t see the way work actually happens inside companies – scattered across disconnected tools, gated by permissions, shaped by undocumented exceptions, and held together by the soft logic of human judgment that never makes it into a system of record.
The big push of 2025 was getting agents to actually do work – not just generate outputs, but execute workflows end to end. This is the “services as software” thesis in action, and it’s where the gap between demo and deployment becomes painfully clear. You can get to 80% with 20% of the effort – enough to close a pilot. But production demands 99% or more, and that last stretch can take 100x more work.
2026 is when startups catch up to the ambition, and when enterprises move from pilots to production. The SaaS incumbents are inadvertently helping: as Salesforce, ServiceNow, and Microsoft push their own agentic offerings, they’re legitimizing the category and making companies more willing to bet on startups who can move faster.
We’ll see startups partner with enterprises to build environments where AI systems can observe, practice, and improve on workflows that don’t exist in any training set. For many tasks, small, customized models running inside enterprise infrastructure will outperform frontier models – they’re faster, cheaper, and able to operate where data can’t leave the building. Security and privacy concerns will reinforce this trend: as agents get real privileges inside core systems, on-prem is getting hot again (more on this in prediction #3).
None of this can happen from 30,000 feet. The startups that crack enterprise reliability will embed engineers with customers (accelerating the FDE trend that emerged this year) to surface unwritten rules and iteratively improve agents through 1000s of edge cases that only appear in production. This is painstaking work – and it’s why I believe deeply technical founders will have a decisive edge.
🧠 Prediction 2: Decision traces become the new data moat
My first prediction is about getting agents into the execution path. This one is about what happens once they’re there.
When an agent executes a workflow, it pulls context from multiple systems, applies rules, resolves conflicts, routes exceptions, and acts. Most AI systems discard all of that the moment the task is complete. But if you persist the decision trace – what inputs were gathered, what policies applied, what exceptions were granted, and why – you end up with something enterprises almost never have: a structured, replayable history of how context turned into action.
We call this the context graph: a living record of decision traces stitched across entities and time, so precedent becomes searchable. It explains not just what happened, but why it was allowed to happen. And it compounds. The more workflows you mediate, the more traces you capture. The more traces you capture, the better you get at automating the next edge case. Data is no longer the new oil; it’s decisions – the map of how the organization actually works.
Startups have a structural advantage here. Because they sit in the execution path, they see the full context at decision time. Incumbents are either siloed or in the read path rather than the write path (data warehouses receive information via ETL after decisions are made – by then, the decision context is gone). SaaS incumbents can add AI to their data, but they can’t capture what they never see.
🛡️ Prediction 3: AI-related security issues take center stage
Agents that execute workflows are holding something incredibly sensitive: not just records of what happened, but the logic of how the business actually runs. The threat surface spans the model, the agent, and every system it touches. The more we standardize “plug an agent into everything” via protocols like MCP, the more we create a loose front-end that creates massive exposure if authentication and access controls don’t keep pace.
Agents fundamentally change the shape of risk. Most legacy security frameworks weren’t built for software that can act on its own. This creates an opening for a new generation of security startups and category winners.
In 2026, I expect AI security to become as standard a board-level metric as cybersecurity readiness. Zero-trust principles will be applied to agents: least-privilege access, explicit checks for anything beyond a narrow scope, and real-time behavioral monitoring that flags unusual patterns. This will slow some rollouts, but it will also mature the field. A bank that’s nervous about letting AI touch transactions today might move forward once proven frameworks exist.
I also expect 2026 to bring at least one high-profile AI agent incident that forces the issue. It will likely look like an agent doing legitimate work while quietly exfiltrating sensitive context or taking actions it shouldn’t. The early warning signs came this year: researchers found a flaw in Salesforce that could allow attackers extract customer data through prompt injection, and a separate breach at Mixpanel exposed data tied to OpenAI’s platform.
⚔️ Prediction 4: SaaS incumbents fight back
This year, the big SaaS platforms stopped treating AI as a feature they could add on later. Salesforce is the clearest signal – down to the fact that it’s willing to rebrand itself as “Agentforce.” ServiceNow is pushing in the same direction. And it’s not just the top tier: mid-market SaaS companies are also feeling the pressure.
Part of the urgency is defensive. Incumbents are starting to see churn at the edges. They also see what the new AI-native products are doing: sitting on top of existing systems, pulling context across them, and creating value the platform doesn’t capture. When that value comes from data inside the incumbent’s product – think Glean building on Slack data – it’s easy to understand why the platform’s instinct is “That’s ours.” Even when the customer is the one who ultimately owns the data, the incumbent still controls the pipes.
In 2026, I expect the incumbents to assert that control more aggressively. Some of it will be framed as security and privacy (and sometimes it will be genuinely motivated by those concerns), but the effect will be the same: tighter access, more friction, and more rules. Expect more limits on API usage, more restrictive terms, and more integration hurdles. In parallel, incumbents will keep pushing native assistants into every surface so the default experience is “good enough” without leaving the platform – AI in every Google Workspace document, a helper on every Salesforce screen.
For startups, this makes dependence a real strategic risk. If your product needs another company’s data access or distribution to work, you should assume that access will get harder, not easier. The ecosystem won’t shut down overnight, but it’s likely to get messier and less open for a while.
🛒 Prediction 5: Agents eat e-commerce
Consumer usage of AI is becoming a habit, and e-commerce is where that shift will show up first.
The last era of discovery was built on ad-driven platforms that inferred preference from clicks and demographics. AI can learn preference in a much more direct way. As text and images converge into a single discovery surface, it becomes natural to start a purchase the same way you start everything else now: by describing what you want.
The rails to turn that intent into transactions are already taking shape. Visa declared that 2025 will be “the final year consumers shop and checkout alone.” Mastercard, PayPal, and Google have all launched protocols for AI agents to make purchases on behalf of users.
In 2026, I expect consumer usage to catch up. It will start in the low-stakes, high-frequency corners of spending and expand into anything the AI system can verify. When that happens, it will change the existing power dynamics of the internet.
Brands today optimize to be found by humans – SEO, paid search, influencer marketing, conversion funnels. In an agentic world, they’ll increasingly need to be legible to (and chosen by) agents. A new pay-to-play layer will emerge where placement isn’t just about ranking in search results, but about being surfaced inside an agent’s decision flow – through bidding, commissions on completed transactions, category-level partnerships, etc.
Aggregators will feel the pain first. The value of aggregation collapses when the friction it solves (helping humans make sense of fragmented markets) is automated by agents. Travel is the obvious category (Expedia, Booking), but the same pressure will hit middlemen across financial products, insurance, autos, etc. These businesses were built on winning the click and converting it. When an agent does the research and makes the selection, there’s no human click left to win.
Amazon and Walmart are better positioned because they don’t just win discovery – they control logistics, fulfillment, and returns. If agents become the interface, these incumbents can either build the agent inside their own ecosystems or become the default execution layer external agents plug into. Either way, they stay in the flow.
Marketplaces like eBay land somewhere in between. Agents could make eBay dramatically easier to navigate, especially for long-tail inventory. But they also raise the bar – on listing quality, authentication, and dispute resolution.
Google faces the most interesting tension. For now, AI Overviews are driving growth in overall queries, and search revenue is growing. Google is also actively pushing ads into these AI surfaces (and expanding them). Long-term, the bigger disruption is agentic commerce: if agents start doing the browsing (and learn to route around ads), Google will need to reinvent monetization for an internet where the human search-click-convert funnel is no longer the default path to purchase.
🥊 Prediction 6: Gemini overtakes ChatGPT
In the back half of this year, Google played major catch-up. ChatGPT’s global MAUs grew only ~6% from August to November 2025 (to ~810M), while Gemini’s grew ~30% over the same period. And on the web, Similarweb’s estimates show ChatGPT’s share of AI chatbot traffic falling from 87% to 68% over the past year, while Gemini’s rose from 5% to 18%.
Gemini 3 made this shift impossible to ignore. Two weeks after Google shipped it, OpenAI went “code red” and refocused teams on improving ChatGPT.
That’s why I think 2026 is the year Gemini (and, to a lesser extent, Grok) take real share from OpenAI and Anthropic in consumer usage. Google can put Gemini directly inside Search, Chrome, Workspace, and Android. Grok’s advantage is narrower, but still meaningful: it lives inside a high-frequency consumer surface (X), where trying it is basically frictionless.
Once “good enough” intelligence is everywhere, usage flows to wherever it’s most convenient. This advantage compounds: distribution drives usage, usage generates feedback, feedback improves the product. OpenAI and Anthropic aren’t going anywhere, but the era where one lab feels like the front door to consumer AI is ending. By late 2026, this will look less like a monopoly and more like a three- or four-player race.
💰 Prediction 7: An AI lab goes public
2026 is likely to see at least one – and possibly both – of the leading AI labs file for an IPO. At the level of capex these companies are committing to, public markets will eventually become the only pool of capital big enough to finance what comes next.
Anthropic has hired Wilson Sonsini to prepare for a potential 2026 listing. Its revenue growth has been extraordinary: it started 2025 at a $1B run rate, hit $5B by August and $7B by October. Internal projections targeting $9B by year-end 2025 and as much as $26B in 2026 – making Anthropic one of the fastest-growing technology companies in history.
OpenAI is on a parallel track. It’s expected to hit $20B in annualized revenue this year, up from $3.7B the year before – a 5x increase in 12 months. Reuters reports OpenAI is laying groundwork for an IPO that could value it at $1T, with a filing potentially coming in H2 2026. Over the past year, the company has committed over $1.4T to infrastructure deals with Oracle, Microsoft, Amazon, and CoreWeave. At this scale, even its current $500B valuation and $50B in private funding leaves a significant gap.
If these listings happen, they’ll be among the largest tech IPOs in history, but they could be rocky.
OpenAI has raised enormous sums at sky-high valuations and is burning cash at a staggering rate. The company is projecting cumulative losses of $115B through 2029. Meanwhile, Google has what OpenAI is spending billions to build: custom TPUs, massive distribution, and the internet’s most powerful monetization engine. (I’ve written before about the power of Google’s infrastructure and cash flows.) Google’s marginal cost per query is far less than OpenAI’s. In a price war, Google can sustain losses that would be existential for a standalone lab.
Anthropic faces a different calculus. It has raised far less capital, which lowers the bar for returning money to investors. But its most recent valuation may prove to be a high-water mark – especially if the market cools or growth decelerates.
Going public means GAAP accounting, quarterly scrutiny, and a shareholder base that behaves very differently from private backers. Private investors in frontier AI are long-only believers; they’re buying a vision of the future and they’re patient. Public market investors want to see a near-term path to profitability, not just a path to AGI.
✍️ Prediction 8: Cursor-like interfaces become the default
We’re hitting the ceiling of the chatbot UI. The next wave of value comes from AI that lives inside workflows, not in side panels.
For most users today, AI still lives in a separate window. You gather context from where the work happens, paste it into ChatGPT, write a prompt, wait for a response, then copy the output back into the original environment. In addition to being high friction, this “copy-paste” model cuts the AI off from the context it needs to be maximally useful and breaks the human-AI feedback loop. The user becomes the translator between two systems that should be one.
Cursor figured this out. Its AI has direct access to your codebase: it sees your files, understands your project structure, and can edits code right where it lives. You highlight a block of code, ask for a fix, and it proposes a change as a diff, which you approve or reject with a keystroke. When the code breaks, it sees the error and iterates. The conversation isn’t separate from the work; it’s woven into the surface where the work happens.
In 2026, I believe this pattern spreads beyond coding. “Cursor for X” becomes the default architecture for knowledge work: legal, finance, marketing, sales, operations. These tools won’t feel like chatbots, but they won’t feel like traditional SaaS dashboards either. They’ll blend both modes: open-ended where exploration matters, more constrained where precision matters.
Today, most AI interfaces wait for you to ask. But the best employees don’t work that way: they observe the situation, propose a solution, and ask for sign-off. AI interfaces will catch up to that standard. We’ll stop visiting AI and start approving its suggestions. Intelligence will become embedded in the places where work happens.
3 trends from 2025 I expect to continue
The predictions above are time-bound bets: things I think will happen in 2026. These three trends are broader, structural shifts that are already underway and won’t resolve in a single year. They’re the reason I’m making several of the bets above.
✖️ Trend 1: Scaling laws become multiplicative
For most of the last decade, progress in AI followed a clean rule: more data + more compute + bigger models = better results. Pre-training scaling laws are still real – Gemini 3 is another datapoint that they continue to buy capability. But pre-training is increasingly the floor, not the ceiling. The biggest models are becoming raw material for further differentiation.
We now have three scaling laws, not one. Pre-training remains the most expensive path. Post-training optimization starts with a capable base model and specializes it through fine-tuning, reinforcement learning, and distillation. These techniques aren’t new, but 2025 revealed how much capability they unlock. Test-time compute lets the model spend more cycles at inference: generating intermediate steps, exploring alternatives, verifying its own work.
What matters for 2026 is that these three laws are multiplicative. A stronger base model, post-trained well, given more time to reason, yields compounding gains. And this is also where the capex story gets more nuanced: efficiency and scale aren’t opposing strategies – they’re complements. Frontier labs are both scaling smarter and scaling bigger, investing in optimization while still pushing training clusters and infrastructure spend upward.
✅ Trend 2: Verification is the bottleneck
The fastest progress has been in domains where you can verify the work – coding, math, accounting. The inverse is now one of the hard constraints on autonomy: if you can’t verify a task, you can’t reliably automate it.
That’s why AI’s capability still feels uneven – what researchers call the “jagged frontier.” Models look strong where there’s a clear definition of success, and much shakier where “excellent” depends on human judgment. In those domains you still need a human in the loop – not because the model can’t produce an answer, but because you can’t prove it’s right.
Figure out how to make a subjective task verifiable, and you’ve opened up a new domain AI can own.
📊 Trend 3: Pricing shifts toward outcomes
The era of buying AI because it’s “AI” is ending. As systems move from pilots into production, buyers are becoming more disciplined. In the coming year, they’ll demand ROI, scrutinize pricing, and turn off deployments that can’t defend their spend.
That shift is inevitable once inference costs land on the P&L. In pilot mode, it doesn’t matter if a demo burns GPU hours, as long as it impresses leadership. In production, using frontier models for every task starts getting expensive. Startups will have to adapt, and many already are. Pricing will continue to evolve from activity-based (pay per use) to workflow-based (pay per task), outcome-based (pay per result), and per-agent models that price an “AI employee.”
The deeper point for startups is that pricing strategy is product strategy. If you want to get paid on outcomes, you need instrumentation, attribution, and results reliable enough to stake your revenue on. The startups that can walk into a customer and quantify value will compound. The ones selling AI tools with vague promises will feel budget pressure first.
***
Thank you all for reading and thinking alongside me this past year. See you in 2026!
Published on 12.30.2025
Written by Ashu Garg
