2026 feels like an inflection point. AI’s continued progress implies we’re hitting new kinds of bottlenecks. These bottlenecks also open new windows of opportunities. The constraints are moving down the stack: from compute to memory to storage, and ultimately to human capacity itself. Meanwhile, geopolitics is reshaping who gets to play, markets are pricing in a productivity boom, and the infrastructure layer is getting interesting again.

These are simply patterns we’re watching. Some we’d bet heavily on (predictions 8 and 9). Others are more exploratory (predictions 1 and 6). But all nine feel like they’re pointing at something real about where 2026 is headed.

For those tl;dr:

1. Robotics: 2026 marks the year Robot Foundation Models (RFM) and Real World Models (RWM) transition from research projects to competitive product race.
2. AI Applications: Closed frontier models show scaling diminishing returns; developers increasingly build on open-weight alternatives.
3. AI Infrastructure: AI inference bottleneck shifts from GPU compute to system-level memory and context management; System-level approaches outperform component optimization
4. Long-term Memory: Startups solving AI’s long-term memory problem—catastrophic forgetting, context limitations—emerge with defensible, commercial approaches in 2026.
5. Mercantilism becomes imperialism: US-China competition escalates from trade barriers to territorial control of strategic resources, forming distinct economic and technology bloc
6. Natural Science: China overtakes US in natural science influence driven by sustained research investment while Trump cuts US university funding.
7. Exit Market: Tech Startups with scale and profitability see major IPO exits in 2026 as the “Everything Rally” drives public market appetite.
8. AI Evaluation: AI code generation outpaces human review capacity; AI evaluation tools become critical bottleneck and strategic investment area.
9. AI Productivity Boom: AI productivity gains plus policy tailwinds trigger the “Everything Rally”—broad asset price appreciation led by tech and growth sectors.

1. Robotics: The RFM/RWM Competition Begins

2026 marks the year Robot Foundation Models (RFM) and Real World Models (RWM) transition from research projects to competitive product race.

The pattern feels familiar: it’s the 2022-2023 LLM race all over again, but for physical AI systems instead of text. RFM and RWM development is showing real progress—not just in model quality but in practical approaches to deployment. Google and NVIDIA have already made major framework announcements, and the investment commitments are starting to look serious.

Here’s what makes this different from earlier robotics hype cycles: the data infrastructure is finally catching up. Wookjae recently wrote about how physical AI systems require exponentially more data than language models—LiDAR point clouds, spatial mappings, temporal sequences of how environments change. That was the constraint. But now we’re seeing companies build the pipelines, simulation environments, and evaluation frameworks that may make RFM/RWM training actually feasible at scale.

What we are watching for: RFM/RWM scaling laws papers, competitive benchmarks with leaderboards, Google and NVIDIA developer conference announcements, and robotics startup funding rounds that explicitly position around these frameworks.

How we will see if this is realized: The validation metrics are straightforward. We will start to see published benchmarks (like HuggingFace), developer adoption numbers (downloads, implementations), and real-world deployment case studies.

Investment angle: This is a picks-and-shovels opportunity. The winners will likely be the infrastructure providers building the data pipelines, simulation environments, and evaluation frameworks that every robotics company needs. We may see:

• Data vendor that creates real-world data for RFM/RWM models (ScaleAI for robotics)
• Cheaper sensors that can give richer environmental context to robots at scale (tactile sensors)
• Accurate RWM models
• Benchmark verifiers/providers

2. AI Applications: Open-Weight Models Go Mainstream

Closed frontier models show scaling diminishing returns; developers increasingly build on open-weight alternatives.

The performance gap is narrowing, and the economic incentive is becoming undeniable. Open-weight models eliminate API costs and vendor lock-in at scale, which matters when you’re processing millions of requests. Big Tech and startups are launching experimental applications built on open-weight models—not because they’re ideologically committed to open source, but because the cost-performance tradeoff finally makes sense.

We were skeptical about open-weight competitive viability. The closed model labs have had such a lead in compute, data, and talent that it seemed like open-weight would stay perpetually behind. But here’s what’s changing: the frontier models such as GPT, Claude, and Gemini are running into diminishing returns on scaling, as many AI gurus are pointint out. The improvements since ChatGPT-4 are incremental, not transformational. Meanwhile, open-weight models are catching up fast—80% of the performance at 1% of the cost enables entirely new use cases that weren’t economically viable before.

Another change that we are witnessing in the field is that the general know-hows of fine-tuning open-weight models and building the right kind of tools and workflow for these models have become readily available. Emergence of opencode.ai and oh-my-opencode, and the massive amount of attention they have received already suggest that developers are playing around how to orchestrate different models.

You don’t need frontier model performance for most applications. You need good-enough performance at a price point that makes the unit economics work.

What we’re watching for: Open-weight model download statistics on HuggingFace and Ollama, VC funding into open-weight-based application startups, and enterprise deployment case studies choosing open-weight for production. shifting meaningfully toward open-weight, we’ll know this is real.

How we will see if this is realized:

• the ratio of open-weight vs closed model API calls across the application layer shifting meaningfully toward open-weight
• The market share of LLM providers become increasingly fragmented

Investment angle:

• Focus on application layer startups with defensible moats—vertical specialization, proprietary data, or unique UX that creates lock-in beyond just the model choice.
• On the infrastructure side, look for fine-tuning platforms, model optimization tools, and deployment infrastructure specifically designed for open-weight models.
• Teams with deep knowledge and/or know-hows of how to optimize token usage and cost allocation across multitude of agents will have an advantage
• Avoid commodity wrapper applications that have no differentiation beyond “we use Llama instead of GPT.”

3. AI Infrastructure: System-Level Inference Wins

AI inference bottleneck shifts from GPU compute to system-level memory and context management; System-level approaches outperform component optimization.

Jensen Huang said it explicitly:

“We need more memory, we need more context, we need more SSD.”

That’s NVIDIA’s CEO telling us that the inference bottleneck has moved. Training was about raw compute—throw more GPUs at the problem. This certainly solves the problem by scaling out, but scaling up has been more difficult because of challenges in creating ever smaller chips. And with exploding usage of AI agents, conserving context without sacrificing speed is going to be important: how you manage context across HBM, SSD, and distributed storage. It is now a system-level problem that expands beyond silicon level.

Here’s an analogy that makes this concrete: think of AI inference like memory management in an apartment. HBM (high-bandwidth memory) is the stuff you keep within arm’s reach—books on your desk, snacks on the counter. You access it constantly and need it fast. SSD is the closet storage—things you don’t use every day but need to retrieve regularly without going to the basement. Distributed storage is the off-site unit you visit once a month. The question isn’t how much of each you have—it’s how intelligently you move things between layers so you’re never waiting.

A similar success case is Apple. They figured this out early. Their approach to SSD caching extends memory efficiency at the system level, blurring the line between RAM and storage. Android manufacturers can’t replicate it because they have fragmented AP vendors—no one controls the full stack. That integrated, ‘system-level’ approach is becoming the competitive advantage in AI inference.

The evidence is piling up: NVIDIA-Groq acquisition (Groq specializes in decode optimization), SanDisk and Western Digital stock rallies driven by AI storage demand (separate from the HBM memory boom), and inference CapEx becoming a distinct category from training CapEx.

What we’re watching for: Inference datacenter CapEx announcements broken out separately from training, NVIDIA-Groq acquisition confirmation or similar system-level M&A, SSD manufacturer guidance explicitly citing AI inference as a growth driver, and cloud provider offerings around context window optimization.

How we will see if this is realized:

• More vertical integration by key players through acquisition or strategic partnerships
• Key players such as Nvidia moving closer to final customer (software companies)

Investment angle: Primary opportunity is companies offering system-level inference solutions—integrated memory hierarchies optimized for large-scale AI inference. Secondary play is SSD manufacturers with datacenter exposure (this is a separate thesis from DRAM/HBM players). Tertiary is networking infrastructure for distributed computing, like CXL or NVLink alternatives. The timing matters: CXL technology standard is expected 2027-2028, so 2026 is the year to position ahead of that deployment wave.

4. Long-term Memory: Breakthrough Solutions Emerge

Startups solving AI’s long-term memory problem—catastrophic forgetting, context limitations—emerge with defensible, commercial approaches in 2026.

This is the problem everyone knows exists but no one has really solved yet. RAG (retrieval-augmented generation) was the first-generation innovation, but it leads to context window over-consumption—you’re pulling in huge chunks of text every time the model needs to reference something. Compression and chunking cause detail loss, like JPEG artifacts in an image. You lose fidelity.

Think about the simple act of avoiding a hot kettle. We do this instinctively because we have long-term memory of being burned. We don’t need to re-learn “hot = bad” every time we encounter heat. AI systems lack this fundamental capability. They can’t build persistent experiential memory that informs future behavior without consuming massive context windows or losing critical details.

The major AI labs know this is the bottleneck. Sam Altman at OpenAI, Demis Hassabis at Google, and the Anthropic team behind Claude have all publicly identified long-term memory as a critical unsolved problem. Current workarounds are telling: Claude and Gemini insert user generated context into every chat session to simulate memory. There are open-source tools such as claude-mem that reduce token usage when browsing past conversation history. But this is simply inserting curated snippets, not true episodic memory. The fundamental context window and its lack of long-term memory persists.

Meta’s acquisition of Manus validates that long-term memory solutions have immediate commercial value. Manus successfully sold “virtual machines” that have delivered reliable complex, multi-step task completion without losing context. While this was so impressive, there are now open-source repository that mimic this to a similar level for free. If that’s enough to get acquired by Meta, there’s a much bigger opportunity for solutions that can expand and optimize agent operations.

What we’re watching for: Long-term memory becoming explicit public discourse (conferences, papers, blog posts), GitHub repository activity spikes on memory management approaches, startup funding rounds explicitly positioning around long-term memory, and major AI labs publishing episodic memory breakthroughs.

How we will see if this is realized:

• Crazy seed-round fundings on memory management startups by ex-openAI engineer

Investment angle:

• A completely new architecture which solves memory problem
• Real-time, streamed fine-tuning for AI

High conviction on companies with defensible business models—not just open-source repos. Look for proprietary data structures or system-level integration that creates lock-in. Focus on solutions demonstrating accuracy without compression artifacts or hallucination issues. This is a picks-and-shovels opportunity: every AI company needs this, and whoever wins becomes infrastructure-layer standard. The timeline is tight—MIT labs and open-source communities are proliferating experimental approaches, so the window for commercial defensibility is 12-18 months before best practices commoditize.

5. Geopolitics: Mercantilism Becomes Imperialism

US-China competition escalates from trade barriers to territorial control of strategic resources, forming distinct economic and technology blocs.

The rhetoric is shifting in ways that feel ominously familiar. In the past, we saw rise of mercantilism followed by colonialism, then imperialism. Mercantilism was 2020-2025: tariffs, export controls, supply chain reshoring. We may see colonialism (Venezuela overtaken by the U.S.?), and shortly followed by imperialism is 2026+: territorial control, resource extraction, sphere-of-influence battles.

Trump’s references to Greenland aren’t jokes—they’re leading indicators. Greenland sits on rare earth deposits and controls Arctic shipping routes that become viable as ice melts. The Arctic is the shortest route between Asia and Europe, cutting weeks off traditional shipping lanes. Control of that geography means control of 21st-century trade routes and access to resources critical for technology manufacturing. For U.S. this is a reasonable strategic move as one of the strongest card China can play is restricting rare earth mineral exports. Getting access to Greenland can remove this leverage.

This goes beyond US-China. It’s regional hegemony battles over peripheral nations. Countries are being forced to choose: align with the US technology and economic bloc, or align with China’s. The middle ground is disappearing. We’re seeing economic bloc formation that restricts technology transfers, capital flows, and even talent mobility between spheres.

The pattern echoes 19th-20th century imperialism—Great Power competition over territory and resources, justified by national security and economic necessity. The difference is that instead of competing for coal and steel, it’s rare earths and semiconductor materials. Instead of colonial administration, it’s economic dependency and technological stack lock-in.

What we are watching for: US military or economic actions regarding Greenland and Arctic access, rare earth supply chain policy announcements, non-NATO military alliance formations , and technology export controls expanding beyond chips, rare earth to broader categories. , rare earth production capacity investments outside China, and cross-border investment flows between blocs decreasing.

How we will see if this is realized:

• US “charm offensive” on Greenland resident
• US military presence expansion in Greenland and Arctic regions
• US administration of Venezuela
• China imposing sanctions against Taiwan - US starting military action in

Investment angle: Defense sector rearmament in preparation of post-NATO disintegration. Resource self-sufficiency investments in rare earth mining, processing, and recycling within the US bloc. Companies with supply chains entirely within a single bloc, avoiding cross-bloc dependencies that create geopolitical risk. And longer-term, Arctic infrastructure—Northern Sea Route shipping, logistics, port development as that becomes economically viable.

I’m not saying this is inevitable, but I’m watching to see if this pattern emerges. If it does, the investment landscape changes fundamentally—globalization reverses, and regional bloc affiliation becomes a first-order consideration in every investment decision.

6. Natural Science: China’s Influence Expands vs US

China overtakes US in natural science influence driven by sustained research investment while Trump cuts US university funding.

This is the prediction we are least sure how to position from an investment standpoint, but the pattern feels significant for understanding where breakthrough science originates over the next decade.

Here’s the data point that caught our attention: China has received only 1 Nobel Prize in the sciences over the last 10 years, despite matching or exceeding the US in research paper volume. That Nobel lag reflects a 10-20 year delay—Nobel Prizes recognize work done decades ago. But China’s current research output could show up in future awards, and the trajectory is clear. However, so far, a lot of China’s research papers were in application level, not in basic sciences.

The policy contrast is stark. Trump’s administration is holding US university research funding prisoner for political gains, while China maintains consistent science and technology investment regardless of political cycles. That stability matters for long-term research programs that require decades of continuous funding. The US approach has become increasingly politicized and volatile—funding priorities shift with each administration, making it hard to sustain multi-decade projects.

We expect to see Chinese researchers increasingly dominate high-impact publications in Nature, Science, Cell, and Physical Review. Citation metrics and h-index rankings will shift. More importantly, Chinese researchers will start appearing in Nobel Prize speculation and shortlists, and within 5-10 years, the Nobel distribution will reflect that shift.

What I’m watching for: China vs US publication counts in top-tier journals, citation metrics of leading researchers, Nobel Prize nominations and Lasker Award patterns (Lasker often predicts future Nobels), and international collaboration patterns—which scientists are choosing Chinese vs US institutional affiliations.

How we will see if this is realized:

• Number of research papers produced in basic sciences start to overtake the U.S.
• Breakthrough in moonshot projects announced by Chinese government, such as quantum computing and space exploration (funded and backed by state institutes)

Investment angle: We’re still working on this. Potentially Chinese biotech, materials science, and quantum computing companies gain talent advantages as breakthrough researchers concentrate there. In the US, this could accelerate private research funding and corporate R&D taking a larger role than university-based fundamental research. But I’ll admit, this one feels more like a geopolitical and talent flow observation than a clear investment thesis. The implications are long-term and diffuse.

7. Exit Market: Pre-AI Companies See Major Exits

Tech Startups with scale and profitability see major IPO exits in 2026 as the “Everything Rally” drives public market appetite.

This is probably the easiest prediction to make—it’s straightforward tech cycle pattern recognition, not a novel thesis. Late-stage companies that grew to scale during 2020-2023 are hitting revenue and profitability milestones that make them IPO-ready. The timing aligns with the Everything Rally expectations (more on that in Prediction 9). Public market appetite for tech exposure is recovering after the 2022 correction, and these companies represent the last generation of “pre-AI” business models before generative AI reshape business strategies.

There’s pent-up demand for liquidity. Employees sitting on illiquid equity for 5-7 years want exits. VCs sitting on unrealized gains need distributions. Crossover investors who bet on late-stage companies pre-IPO want validation. And public market investors want access to tech growth that isn’t just the Magnificent 7.

The boom cycle we’re in creates the conditions: sustained tech sector strength, Everything Rally macro backdrop, and NASDAQ appetite for new listings. Late-stage valuations are closing the gap with public market comparables, making IPO windows attractive.

What we are watching for: S-1 filings and IPO announcements from recognizable late-stage companies, NASDAQ new listing pipeline growth, and late-stage private valuations converging toward public comparables. Validation metrics are simple—number of tech IPOs completed in 2026, first-day and 30-day performance, and NASDAQ outperformance vs S&P 500 confirming tech-led rally.

How we will see if this is realized:

• Decade-old unicorns such as Stripe, SpaceX and Databricks, Discord file for IPO
• The stock prices of these players end at 100% over their initial IPO market cap in 12 months
• Top-tier venture capital firms are increasingly retaining their positions post-IPO

Investment angle: Pre-IPO exposure through late-stage secondaries and pre-IPO funds, if you have access. Public market participation in IPOs or early trading days. Focus on companies with clear paths to profitability, not pure growth stories—public markets in 2026 will reward sustainable unit economics over “grow at all costs.”

8. AI Evaluation: Humans Become the Bottleneck

AI code generation outpaces human review capacity; AI evaluation tools become critical bottleneck and strategic investment area.

“1 engineer, 1 month, 1 million lines of code” is the new standard of software engineer productivity according to Galen Hunt at Microsoft. While AI can definitely give you a million lines of code, one can’t carefully read, test, and validate that volume. Something has to give.

Satya Nadella entered what people are calling “founder mode”—communicating directly with AI developers, delegating legacy business operations to focus on AI development. When a CEO of Microsoft’s scale does that, it’s a signal that this is considered a do-or-die issue..

AI sped up software engineering at first, and this will expand to other jobs and industries. The bottleneck shows up everywhere: code review backlogs, QA cycle times, security audit delays, and pull request merges taking days instead of hours. Current solutions are DIY internal tools—Cursor’s debug mode, Claude’s markdown session tracking, custom Git branch management scripts. These are stopgaps, not solutions.

There’s a precedent here: Airbnb developed Airflow internally for data pipeline management, then open-sourced it to Apache. It became infrastructure for the entire ecosystem. I keep wondering which company creates the Airflow equivalent for AI code evaluation—the framework that becomes the standard for managing AI-generated code at scale. And the know-how of implementing this will be an edge for enterprises of tomorrow.

The problem is simultaneously obvious and hard to get right. Everyone’s building internal AI evaluation tools. The question is which one escapes internal use and becomes the commercial winner. Do you bet on open-source frameworks gaining traction (the Jenkins model), or do you bet on enterprise commercial solutions with better UX and support (the GitHub Actions / CircleCI model)? And who will make AI evaluation for other tasks such as marketing, sales and finance?

What I’m watching for: AI evaluation tool startups securing funding, Big Tech acquisitions of evaluation and testing companies, open-source frameworks gaining contributor momentum, and enterprise surveys explicitly identifying developer productivity bottlenecks. Validation metrics include adoption numbers (MAUs, seats sold), code review cycle time improvements, reduction in post-deployment bugs, and developer satisfaction scores.

How we will see if this is realized:

• AI evaluation becomes a standardized process in software engineering, much like QA
• Framework of AI-driven reviews and how to ‘review’ the reviews become a popular subject
• AI evaluation tools secure funding

Investment angle: High priority on AI evaluation tools with enterprise traction. Look for solutions handling the full stack: code review, security scanning, integration testing, and architecture analysis. Avoid pure open-source plays without defensible business models—history shows that open-source infrastructure gets monetized by adjacent commercial products, not the core project itself. Compare to CI/CD market structure for pattern recognition.

Every company using AI code generation likely hits this bottleneck. The TAM is every software company globally. The question isn’t whether the market exists—it’s who wins it.

9. AI-Driven Productivity Boom: The Everything Rally

AI productivity gains plus policy tailwinds trigger the “Everything Rally”—broad asset price appreciation led by tech and growth sectors.

This is the most macro of the nine predictions. Whether the others come true or not, this one depends on policy choices that feel increasingly likely.

This thesis has three components that need to align:

Productivity: We have seen countless newspaper articles and case studies that tout crazy productivity increase for AI code generation, AI customer service, AI content creation. Macro data is still somewhat ambiguous, but the policy intent seems clear.

Policy: Kevin Hassett, potentially Trump’s pick for Fed chair, has advocated for rates as low as 1% (compared to the current 3.25%). The logic echoes Alan Greenspan’s 1990s-2000s playbook: if productivity is rising, you can run the economy hotter without triggering inflation. The Phillips curve relationship between unemployment and inflation breaks down when productivity accelerates. Greenspan argued this during the internet boom, and Hassett appears positioned to make the same argument for the AI boom.

Fiscal: Trump’s “One Big Beautiful Bill”—tariffs up, taxes down, spending up. Fiscal stimulus plus monetary easing plus productivity gains create the conditions for an Everything Rally. Not just stocks—bonds, real estate, crypto, everything rises together because the macro environment supports broad asset price appreciation.

If Hassett becomes Fed chair, we are watching for an immediate 5+% NASDAQ jump as confirmation that markets are pricing in this policy shift. That’s the leading indicator. From there, validation metrics are straightforward: all asset classes rising, NASDAQ outperforming S&P 500 (tech-led rally), volatility staying low despite rate cuts (confidence signal), and credit spreads tightening (risk appetite confirmation).

What I’m watching for: The critical signal is the Kevin Hassett Fed chair appointment. Secondary indicators include tech sector employment data (if Silicon Valley stays stable despite AI deployment, it validates that productivity gains exceed displacement effects), actual rate cut trajectory, and Total Factor Productivity data continuing upward trends.

How we will see if this is realized:

• Productivity growth accelerates past 2025 2H, showing consistent acceleration for the next 8 quarters
• Fed target range goes down by 1%p by year-end, 1%p lower than the current outlook
• Inflation remains stable at 2~% even with market rally

Investment angle: Highest conviction is leveraged tech exposure if the Hassett appointment happens. Sector focus on Big Tech and established AI infrastructure players—the “pinnacle” companies with pricing power and scale advantages. Risk management: Everything Rally thesis requires policy plus productivity alignment. If either breaks—if inflation resurges or productivity gains don’t materialize—the thesis fails.

The pessimist in me wonders if this creates unsustainable asset bubbles that unwind badly in 2027-2028. But for 2026 specifically, the setup looks pretty clear.

Where I’m landing with this

Several of these predictions interconnect in ways that feel significant. Predictions 3, 4, and 8 all point to the same bottleneck shift: compute → memory/context → human evaluation. The constraint is moving down the stack and eventually hitting human capacity itself. Predictions 1, 2, and 7 reflect a maturation cycle: foundation models → applications → exits. And Predictions 5 and 6 show geopolitical forces reshaping who gets to participate in the technology development cycle.

Prediction 9—the Everything Rally—provides the macro context that amplifies all the others. If asset prices surge broadly, every thesis becomes more possible. Funding flows increase, exit markets open, risk appetite expands. The macro tailwind makes infrastructure bets, application layer experiments, and even speculative robotics investments all more viable.

Some are higher conviction than others. I’d bet heavily on 8 and 9. We’re less certain on 6 - we admit that one might be more about geopolitical observation than actionable investment thesis. But these are the patterns we’re watching as we move through 2026.

Here are the open questions I keep coming back to: If humans become the bottleneck (Prediction 8), does that accelerate demand for long-term memory solutions (Prediction 4) or system-level inference optimization (Prediction 3)? Does the Everything Rally survive if geopolitical tensions escalate to actual conflict? Which happens first—RFM/RWM competition or AI evaluation maturity?

What did I miss? What breaks these theses? I’d love to hear what you think. Let me know in the comments or reach out. Let’s talk.