Every year, we publishes our predictions for the year ahead. This is our 2026 edition: 20 predictions across models, products, infrastructure, and markets.

2025 opened with DeepSeek and closed with Manus. In between, we witnessed agentic capabilities make a generational leap, AI coding go from interesting demo to reliable ARR engine, and Google stage one of the most dramatic narrative reversals in recent tech history. The pace of change is accelerating, not stabilizing.

What comes next is the question we've spent months working through — with researchers, builders, and investors across the AI ecosystem. We'll be wrong on some of these predictions. That's the point. Predictions that could plausibly be false are the only ones worth making.

Part I: The Next Paradigm Is Already Being Built

Prediction #1: Continual learning signal emerges — AI begins to learn autonomously

After scaling laws and reasoning models, the field needs a new paradigm. We've believed in continual learning as that next breakthrough since last year — and, to be direct, underestimated how hard it would be. We're not revising the thesis, just the timeline.

Full, general continual learning won't arrive in 2026. But the early signal will. Concretely, we expect the field to converge on 1–2 viable technical paths, producing work that functions like Transformers did in 2017 or Scaling Laws did in 2020 — not yet dominant, but clearly the direction. Models today struggle to create decisive generational gaps between iterations. A new paradigm is not just desirable; it's the only path to a genuine step-change in capability. The density of talent now focused on this problem — OpenAI, SSI, Thinking Machines Lab, and others — makes a null result increasingly unlikely.

Prediction #2: World Model paths converge — entering the GPT-2 moment

World Models are closely related to multimodal progress, and in 2026 we expect the exploratory phase to end. Two main camps have been developing in parallel: one focused on real-time video generation for consumer entertainment and games, the other on physical accuracy for robotics and autonomous driving. The "blind exploration" era closes as a technical consensus forms — what we're calling the GPT-2 moment for world models. After that inflection, the work shifts from "which direction" to "how far can we scale." The real explosion comes in 2027 and beyond, but 2026 is when the foundation gets laid.

Part II: ChatGPT Becomes the Internet's New Front Door

Prediction #3: ChatGPT DAU doubles — reaching true global-gateway scale

ChatGPT's current DAU sits around 400–500M. We predict this doubles to 800M–1B in 2026. The 1B DAU threshold matters because it marks the point where ChatGPT stops being a tool and starts functioning as infrastructure — the kind of position that lets a platform extract value from the economy it sits above, rather than merely participating in it.

The web traffic signal reinforces this framing. Today, Google and ChatGPT split search-adjacent traffic roughly 85:15. We expect that to shift to at least 70:30 by end of 2026. Established players across industries — having navigated the disruptions of Web 1.0 and 2.0 — will start moving from "wait and see" to active partnership. Disney and Intuit have already taken early steps. The move accelerates as scale becomes undeniable.

Prediction #4: ChatGPT's "App Store Moment" — the first $100M ARR native app

We've long believed LLMs are the new OS. The App Store Moment is when that analogy stops being theoretical: ChatGPT transitions from super-tool to super-platform, producing a killer native app that generates $100M ARR entirely within the conversation interface — users never need to leave the chat. Unlike the mobile App Store's list-and-browse model, we expect this to be AI-recommended and contextually embedded. The app doesn't get discovered — it gets suggested at exactly the moment of relevance. ChatGPT's expanding ad ambitions make app recommendations a natural commercial layer, reinforcing the platform flywheel.

Prediction #5: OpenAI narrative reversal in H2 2026 — valuation breaks $1T

OpenAI's current position resembles Google's in early 2025: under real competitive pressure, narrative softening. But two catalysts could flip the story by H2. First, if ChatGPT crosses 1B DAU, it formally becomes a "productivity toll collector" — the kind of structural leverage that triggers sharp re-rating. Second, OpenAI's institutional DNA and technical depth make it arguably the best-positioned of the Top-3 labs to produce the next paradigm signal. A continual learning breakthrough in H2 would restore clear SOTA status and trigger at least a half-year of clear leadership. With both catalysts in play, a $1T valuation — currently contrarian — becomes defensible math. And given NVIDIA's deep narrative entanglement with OpenAI, a re-rating likely pulls NVIDIA along with it.

Prediction #6: Personalization becomes the moat — Proactive Agents lock in users

The playbook was proven by ByteDance: learn user data deeply → deliver personalized experience → raise switching costs until migration becomes irrational. In the AI era, the same dynamic plays out through context and memory accumulation. ChatGPT's Pulse feature points in this direction — moving from passive search tool to proactive task planner that pushes relevant content based on a deep model of who you are and what you're working on.

The strategic implication is significant. As models evolve from "general tools" to "dedicated agents," the competitive moat shifts from benchmark performance to accumulated user context. A model that knows your communication style, working patterns, ongoing projects, and long-term goals has a structural advantage that a technically superior but contextually naive competitor cannot easily overcome. This is not a soft moat. It compounds.

Another key insight is that Gemini, GPT, and Claude are unlikely to achieve a disruptive technical lead over each other at the model level. With pre-training and reinforcement learning entering the industrialized era, the real focus shifts to strategic bets — again, the keyword is differentiation. It's entirely possible that these three models will evolve very differently in practice, much like the e-commerce landscape has multiple distinct players.

For example, ChatGPT is shaping up as a personal assistant, while Gemini remains more of a "workhorse" tool optimized for domains like healthcare and medical problem-solving.

Part III: Enterprise AI — The Year It Actually Ships

Prediction #7: 2026 is the Enterprise AI breakout year — Anthropic ARR at least doubles

The 2025 agent progress made AI-created economic value a real phenomenon. In 2026, that potential gets realized at enterprise scale. Two patterns will dominate: Buy (prosumer-facing, out-of-the-box products displacing the MS Office stack — Google's tools are already pressuring PowerPoint, and Anthropic is building around spreadsheets) and Build (enterprises constructing agentic workflows directly on model APIs).

Anthropic is structurally positioned to be the biggest winner on the Build side — not because of absolute model superiority, but structural position. OpenAI is tightening its Microsoft alignment and pivoting toward ChatGPT Enterprise as its primary enterprise vehicle. Google's Gemini API remains tightly coupled to GCP, limiting ecosystem portability. Anthropic is the only Top-3 lab operating as a genuinely neutral, multi-cloud API provider — a position that becomes increasingly valuable as enterprises want frontier capability without infrastructure lock-in. We expect ARR to at least double, potentially surpassing $20B, with meaningful upside beyond that.

Prediction #8: Long-horizon tasks and multimodal drive a new wave of $1B ARR data companies

Data companies' structural advantage persists — it just migrates with each new capability frontier. Scale AI's breakout came from pre-training data. Mercor and Surge AI rode the RL demand wave. The 2026 driver is twofold: long-horizon task completion requires massive, complex, long-trajectory interaction data that simple Q&A pipelines can't produce; and multimodal creates entirely new labeling and processing dimensions that existing data infrastructure wasn't designed to handle. Against the backdrop of Enterprise AI's expansion, domain-specific proprietary data suppliers also see a significant opportunity uplift — AI delivery quality is directly correlated with domain knowledge depth.

Prediction #9: AI company valuation consensus forms — agent economics get quantified

The standard SaaS metrics — DAU retention, seat counts, Rule of 40, gross margins — were designed for a world where software augmented human workers. They're increasingly inadequate for agent businesses, where the unit of value is a task completed or a decision made, not a seat licensed. The field has been converging on "economic value delivered" as the right conceptual framework, but without agreed-upon quantified standards. As proactive agents and long-horizon tasks become product realities — with Cursor, ChatGPT, and Glean as reference cases — we expect a new evaluation framework to emerge in 2026 that the investment community coalesces around. This isn't just academic; it's what determines who raises capital and at what multiple.

Part IV: The Global Model Race Splits

Prediction #10: xAI merges into Tesla — bridging digital and physical AGI

xAI fell behind in 2025's model competition. A Tesla merger is the move that converts the liability into a strategic asset: xAI's intelligence infrastructure becomes Tesla's embodied AI premium. The real thesis was never whether Grok would become the next ChatGPT. From day one, xAI's most interesting scenario was connecting digital intelligence to physical systems — FSD, Optimus — in a way no pure software lab can replicate. As AI moves deeper into multimodal and World Models, that physical-digital integration thesis becomes more compelling, not less. The 1+1+1 > 3 logic is strongest precisely at the frontier where it's hardest to execute.

Prediction #11: Chinese companies continue to dominate open-source SOTA — the US has no equivalent

Starting with DeepSeek's January 2025 breakthrough, the open-source SOTA crown shifted from the US to China. We expect that shift to become permanent in 2026. Qwen and DeepSeek will continue holding global Tier-1 open-source positions, with global Tier-1 open-source entirely China-occupied.

The structural explanation is underappreciated in Western AI discourse: China's top labs are deploying their best researchers on open-source as a market-share and influence strategy. Meanwhile, OpenAI and Anthropic concentrate resources on proprietary systems to protect commercial moats. This isn't a failure of US capability — it's a rational commercial choice that leaves a competitive vacuum. The constraint became the strategy. Chinese companies' more challenging commercialization environment pushed them toward open-source as a "land grab" for developer mindshare, and that pressure produced dominance.

Prediction #12: Google breaks $5T market cap — M7 divergence accelerates

Google's comeback won't stall at 2025 levels. We expect the market cap to push past $5T, implying EPS of $16–20 and a 30x+ PE multiple — a valuation that requires believing in Google's durable AI advantages. The conviction comes from three angles: structural multimodal depth that makes Google the clearest beneficiary of the multimodal breakout year we're forecasting; an advertising business more resilient to AI disruption than consensus assumes, and potentially growing with AI as an incremental demand driver; and valuation anchoring — if OpenAI reaches $1–1.5T, Google at $4–5T is defensible math.

The broader M7 picture splits into three tiers. NVIDIA and Google are the structural winners. Apple and Tesla hold middle ground through hardware and physical AI positioning. MSFT, AWS, and Meta are likely to lag — their AI bets have been slower to compound into distinct competitive advantages, and the market will increasingly price that differentiation.

Part V: Infrastructure — The Real Long-Duration Bet

Prediction #13: Inference compute explodes, token consumption grows at least 20x

2025 showed the direction. 2026 is the detonation. Hardware capacity is capped by TSMC throughput, but token consumption has its own logic. Three structural drivers: the shift from instant Q&A to multi-day long-horizon tasks (one complex task = hundreds to thousands of inference calls, compared to one for a simple question); proactive agents generating inference without human triggers, creating continuous background compute demand; and multimodal and World Model workloads that require sustained spatial, video, and long-trajectory compute far beyond text-only inference. The delivery of NVIDIA's Blackwell GB series will improve inference economics and further stimulate consumption — a supply-side enabler meeting explosive demand-side growth.

Prediction #14: NVIDIA aggressively enters co-packaged optics — CPO M&A wave begins

The AI compute bottleneck has migrated from calculation speed to data throughput. Co-packaged optics (CPO) is the answer at the 3.2T interconnect generation, where conventional electrical interconnect approaches hard physical limits. NVIDIA's micro-ring modulator technology could allow it to leapfrog Broadcom's current CPO leadership, creating a meaningful technology gap at the interconnect layer. We expect NVIDIA to move aggressively — through acquisition or partnership — toward companies such as Ayar Labs and Lightmatter, triggering a broader CPO M&A wave across the semiconductor ecosystem.

Prediction #15: Multimodal drives a storage revolution — memory integrates into compute core

The 2025 storage rally was cycle recovery. The 2026 driver is structural transformation. Multimodal training and inference shifts hardware attention from pure compute to storage bandwidth and capacity — particularly eSSD, where the bandwidth-per-watt profile suits sustained video and spatial workloads. If World Model paths converge (Prediction #2), tiered storage demand cascades through the entire stack from NAND to high-performance controllers. NVIDIA and Google are likely to design dedicated video ingestion chiplets in 2026, marking storage's formal integration into the compute core. HBF samples are expected by year-end, with actual delivery in 2027–28 — creating a valuation-premium window that will almost certainly overshoot near-term fundamentals before correcting.

Prediction #16: Autonomous driving enters commercial scale-up — Tesla Robotaxi and FSD hit inflection

2016 was when the industry started betting seriously on autonomy at scale. 2026 is the 10-year mark, and we believe it's the year the transition from "experimental technology" to "large-scale commercialization" becomes official. FSD subscriptions are at 700K users (roughly $700M annualized at $1,000/year). We expect that to approach 1.5M by end of 2026 — becoming one of Tesla's most important cash flow lines. Global Robotaxi fleet volume could grow from ~20K to ~300K vehicles. Beyond Tesla, Waymo (~3,000 vehicles currently), Pony.ai, Baidu, DiDi, and WeRide are all at the critical inflection from "thousands" to "tens of thousands" — the threshold where unit economics start to genuinely compound.

Part VI: The Wilder Calls

Prediction #17: Multimodal hits its "AI Coding moment," and an AI Pokémon GO emerges

AI coding (Claude Code, Cursor, Lovable) proved that a technical capability can abruptly start generating scalable commercial revenue. We expect multimodal to hit the same inflection in 2026, producing multiple companies with real, growing ARR across AI-native infra, terminal agents, and consumer products. The distinctive characteristic of multimodal versus coding: far stronger consumer applicability. Pokémon GO pulled AR out of the lab and onto streets worldwide for hundreds of millions of people — while accelerating smartphone hardware cycles. We expect a multimodal product to do something analogous for AI hardware in 2026. The software-hardware co-evolution that Pokémon GO catalyzed for smartphones is coming for AI devices.

Prediction #18: Agentic Web triggers platform wars — Apple enters with a decisive protocol

ByteDance's Doubao Phone crystallized a tension that was always coming: AI agents operating across apps and services fundamentally threaten every platform's traffic-distribution model. The historical analog is the "3Q War" of 2010, where 360 Security attempted to interpose itself between users and QQ by acting as a cross-platform agent — triggering an existential conflict between Chinese internet platforms. The agentic version is structurally identical but orders of magnitude larger in scope. We don't expect any Android OEM or standalone AI company to successfully navigate this before Apple acts. When Apple does — leveraging its full-stack hardware/software/ecosystem position to impose Agentic Web protocols — software vendors will yield partial control to stay in the iOS ecosystem. That's Apple's 2026 AI "comeback" play.

Prediction #19: An AI scientist solves a major scientific problem

The 2025 signals were clear: Terence Tao and collaborators using AI for mathematical discovery, research agents entering literature review and experimental design. In 2026, AI's scientific participation deepens — not assistance to human scientists, but primary authorship of a significant result. The most likely domains are mathematics, physics, or materials science, where formal verification and structured reasoning give current models the most reliable foothold. OpenAI, Google, and others are actively scaling their science-focused investments, and the competitive pressure to demonstrate scientific impact is real.

Prediction #20: AI IPO boom, then a sentiment peak

Zhipu and MiniMax going public marks the opening act. SpaceX, OpenAI, and Anthropic have all floated IPO timelines. Vertical AI companies with proven revenue — Abridge, ElevenLabs, and others — complete the wave. A year with this concentration of high-profile listings implies a significant bull market: elevated sentiment, intense retail interest, and extreme capital market participation.

Historical precedent is equally clear: clustered mega-IPOs mark sentiment peaks. The macro confluence to watch — end of the Fed rate-cut cycle, inflation re-emergence, IPO capital absorption — could turn the rally into a meaningful correction by H2. We're constructive on the first half; cautious about extrapolating it through year-end.