Executive Summary
AI investing has often been framed as a choice between “models” and “applications” over the past two years. This framework was useful in the early phase of the cycle, when model capability was the scarcest asset. The companies that could train the strongest models stood closest to developer adoption, enterprise budgets, capital formation, and ecosystem control.
However, in 2026, this binary had lost much of its explanatory power.
AI is becoming a full industrial system spanning power, chips, data centers, models, agent runtime, workflows, payments, identity, and service distribution. Many companies now operate across multiple layers: one company may sell chips and cloud, train models and provide tooling, or run an exchange while trying to define payment standards for agents. Meanwhile, the “application layer” is splitting into distinct asset classes: vertical agents, harness middleware, agent-commerce protocols, and AI-enabled service rollups.
That is why Foresight Ventures believes AI investing needs to move beyond the "models vs. applications" binary and into a layered framework. What ultimately shapes returns is not simply whether a company sits in the model layer or the application layer, but what kind of moat it actually owns: physical scarcity, capital and talent oligopoly, execution-trajectory data, customer and workflow data, or network effects and standards lock-in.
NVIDIA’s “AI five-layer cake,” presented at GTC 2026, is an important starting point. It divides the AI industry into energy, chips, infrastructure, models, and applications, reminding the market that AI is an industrial-scale value chain rather than a single software innovation. But NVIDIA’s framework reflects a hardware platform’s perspective. It helps explain how NVIDIA captures value across layers, while leaving open the venture question: where can new companies still be built?
Foresight’s framework further divides what NVIDIA compresses into the application layer. We map AI investing into five layers:
- Physical Substrate
- Frontier Intelligence
- Harness
- Workflow Ownership
- Agent Commerce + Identity
Foresight Ventures’ five-layer AI investment map
Our core view: AI applications should be judged by workflow control, not by whether they “have an agent.” The most valuable AI companies of the next cycle will enter real businesses, control execution chains, accumulate outcome data, and turn AI transformation into better unit economics.
This is also where the crypto and exchange opportunity sits. Agent commerce, identity, payments, settlement, and service discovery will become part of the economic substrate of the agent economy. However, core standards positions are being captured quickly by Coinbase, Cloudflare, Stripe, Visa, Google, Anthropic, and other platforms. For Foresight and Bitget, the practical opportunity is to secure early positions in cross-rail settlement, KYA, agent identity, trading-execution harnesses, and high-value financial workflows.
Why “Models vs. Applications” Is Breaking Down
At the start of the AI cycle, it was rational for the market to focus on models itself. Without strong foundation models, there would be no coding agents, AI customer support, AI search, agentic workflows, or automated trading. Model capability sets the ceiling of the entire ecosystem, which explains the premium valuations of OpenAI, Anthropic, Google DeepMind, xAI, and a small number of leading model companies.
As model capability diffuses, the investment question changes. Enterprises care whether AI can connect to internal systems, obtain permissions, execute long-running tasks, remain accountable, expose its process for audit, measure outcomes, and improve after failure.
AI value chain testing who can turn intelligence into execution. The old application bucket now contains at least three different opportunity sets:
- Harness: memory, tool calling, orchestration, evaluation, tracing, and guardrails.
- Workflow Ownership: business processes in customer service, legal, accounting, insurance, recruiting, financial research, and other verticals.
- Agent Commerce + Identity: how agents discover services, authenticate, obtain permissions, pay, and establish trust.
These categories have different moats, customers, business models, and exit paths. Underwriting them as one “application layer” leads to weak pattern recognition.
Foresight’s starting question is therefore broader: which layer of the AI value chain does this company occupy, and what durable barrier does it own?
From NVIDIA’s Hardware Stack to Foresight’s Venture Stack
NVIDIA’s “AI five-layer cake” divides the AI system into energy, chips, infrastructure, models, and applications. Its key contribution is to frame AI as an industrial system.
From NVIDIA’s perspective, the first three layers form the AI Factory. Power, chips, data centers, networking, cooling, and cloud infrastructure determine the productive capacity of the AI era. As models and applications grow, demand for the AI Factory increases. NVIDIA’s strength comes from operating across chips, networking, CUDA, NIM, DGX Cloud, robotics, and autonomous driving, anchoring more of the value chain around its platform.
Venture investors need a different map. The central questions are: which layers can still produce new companies, which moats are durable, which entry prices are sensible, and which paths to exit are realistic?
Foresight reconstructs the AI investment stack as follows:
| Layer | Name | Moat Type | Investment Logic |
| 1 | Physical Substrate | Capex + physical scarcity | Back the load-bearing pillars of AI demand rather than direct combat in silicon |
| 2 | Frontier Intelligence | Capital + talent oligopoly | General models are concentrated; focus selectively on Physical AI and World Models |
| 3 | Harness | Trajectory data | Standalone middleware is compressing; prioritize vertical integration or M&A optionality |
| 4 | Workflow Ownership | Customer + workflow data | Focus on who owns the customer, process, outcome data, and pricing power |
| 5 | Agent Commerce + Identity | Network effects + standards lock-in | Track payments, identity, service discovery, connectors, and settlement standards |
The point of this framework is that it is a reminder that every layer requires a different due-diligence method. Physical Substrate requires analysis of capacity, orders, PPAs, power costs, and geopolitical risk. Frontier Intelligence requires judgment on talent density, capital intensity, and data flywheels. Harness requires analysis of execution trajectories and the optionality of strategic acquisition. Workflow Ownership requires evaluation of customers, processes, gross-margin improvement, and outcome-based pricing. Agent Commerce requires attention to protocol adoption, standards lock-in, and cross-organization call volume.
Layer 1: Physical Substrate — Back the Load-Bearing Pillars of AI Demand
Physical Substrate is the base layer of AI: chips, packaging, memory, networking, optical communications, power, cooling, data centers, and neocloud. This layer resembles semiconductor, energy, and industrial research more than software investing. It speaks through capacity, orders, capex cycles, gross margins, customer concentration, supply chains, and geopolitical exposure.
In Foresight’s perspective, the most compelling areas are power, cooling, networking, packaging, and leading neocloud players. Independent AI-chip startups are harder to underwrite.
The key new variable is agentic workload intensity. Agents turn a single interaction into a long-running, multi-step execution chain. One task may include planning, retrieval, tool use, code execution, retries, permission checks, and result verification. Demand should therefore be modeled through agent step count, retry rate, and tool-call density, rather than simple API-call curves. This amplifies load across compute, memory, networking, and observability.
The independent silicon window is narrowing. NVIDIA, AMD, Google TPU, AWS Trainium, and Microsoft MAIA already form a formidable competitive landscape. A chip startup pitching “cheaper or faster than NVIDIA” must overcome financing risk, software-ecosystem risk, customer-migration risk, and mass-production risk at the same time.
Power and cooling look more like hard bottlenecks. Even as model prices fall and inference efficiency improves, data-center expansion still requires power access, heat dissipation, transformers, power distribution, and construction lead time. Compute competition increasingly becomes energy and infrastructure competition.
Neocloud sits between public and private market exposure. Companies such as CoreWeave, Nebius, Crusoe, and GMI Cloud benefit from GPU demand, outsourced AI workloads, and supply gaps outside major cloud providers. They also carry high customer concentration, heavy capital intensity, and cyclicality risk if supply catches up with demand.
A valuation shift is also underway. Memory, networking, optical-module, and equipment companies were historically priced around book value, inventory cycles, and pricing inflections. If agentic workloads sustain stronger demand, assets such as Micron, Arista, Coherent, and Lumentum may migrate toward growth-oriented valuation frameworks. Investors will ask whether agent workload can support a steeper demand curve over the next three to five years.
NVIDIA’s narrative will also evolve. The first bull case was GPU scarcity. The next question is whether GPUs, memory, networking, and storage can be orchestrated efficiently enough for complex agentic workloads. The bottleneck is moving from GPU availability to AI-factory utilization, throughput, and unit-token cost.
Foresight’s stance: Physical Substrate is investable, but it should be approached through public markets, pre-IPO opportunities, structured deals, or participation in category leaders. Venture should avoid making early hardware startups its main battlefield.
Key diligence items:
- Order visibility
- Locked-in power pricing and PPA structure
- Unit economics per MW
- Customer concentration
- Depreciation cycles
- Supply-chain exposure
- Geopolitical risk
Key disconfirming signal: hyperscaler capex utilization falls materially while agent demand fails to absorb new supply.
Layer 2: Frontier Intelligence — General Models Are Concentrated; Physical AI Still Has a Window
Frontier Intelligence is the capability layer of AI models: general text models, multimodal models, video and audio models, code models, world models, robotics models, and bio-frontier models. This layer is defined by capital and talent concentration. The strongest models require elite researchers, large compute budgets, data flywheels, and distribution channels. General models have become difficult for ordinary early-stage VC to underwrite.
Foresight sees the general-model layer has entered an oligopoly phase. Incremental alpha is more likely in unsettled sub-layers, especially Physical AI and World Models.
The text frontier is already concentrated among OpenAI, Anthropic, Google DeepMind, xAI, Meta, and a small number of leading Chinese model companies. These players control most of the capital, users, developers, and enterprise entry points. A small-model startup may still win on a benchmark, but that advantage rarely compounds into a durable standalone company without distribution, data, and compute access.
Code models are also being absorbed into broader platforms. Claude Code, GPT, Gemini, Cursor, Devin, and related products are collapsing coding-model capability and coding-agent workflow into integrated stacks. That leaves less room for independent coding-model companies.
Physical AI and World Models remain earlier. They require models, embodiment, real-world data, sim-to-real capability, hardware supply chains, task coverage, and deployment environments. The data flywheels around robotics, autonomous driving, industrial automation, and embodied intelligence are still forming.
Companies such as Physical Intelligence, Wayve, Waabi, Skild AI, Figure AI, 1X Technologies, and Sanctuary AI remain important to track.
Foresight’s stance: participate selectively in Physical AI and World Models, while avoiding broad exposure to general-model narratives at elevated prices.
Key diligence questions:
- How many embodied systems are deployed in real environments?
- Is task coverage high-frequency enough to generate data advantage?
- Is the sim-to-real gap narrowing?
- Does the dataset come from real-world operation, simulation, or both?
- Should commercialization begin in enterprise and industrial environments or the consumer home?
- Can unit economics work before hardware costs fall meaningfully?
Layer 3: Harness — The Harness Is the Dataset
Harness is the execution system that brings models into production. It includes memory, context, tool calling, skills, orchestration, guardrails, evaluation, tracing, and observability. If models provide intelligence, Harness provides executability.
The market has produced a long list of agent middleware: memory tools, tool-calling platforms, orchestration frameworks, observability products, evaluation platforms, and safety guardrails. These products solve real problems, but standalone middleware is becoming less independent.
Pressure comes from two directions. First, model companies are turning tool calling, computer use, skills, memory, evaluation, and agent builders into native capabilities. OpenAI, Anthropic, Google, and Microsoft have strong incentives to control the core agent runtime.
Second, vertical agent companies that own customers and workflows will internalize critical parts of the harness. Customer-support agents, legal agents, trading agents, insurance agents, and accounting agents will want control over execution trajectories, customer data, evaluation standards, and risk logic.
The durable harness opportunities have two shapes.
- Vertically integrated harnesses for high-risk industries
High-risk industries, including financial services, trading, insurance, legal, and healthcare, need context, tools, permissions, guardrails, tracing, evaluation, and feedback loops fused into one system. The moat here comes from the closed loop, not from any individual module.
- Strategic acquisition assets
Large security vendors, cloud companies, data platforms, and model providers will keep acquiring companies in evaluation, observability, guardrails, AI security, and workflow runtime. Harness middleware may never become the next independent platform giant, but it can still become a valuable strategic asset.
The central idea of this layer is simple: the harness is the dataset.
Foresight’s stance: Whoever captures high-quality execution trajectories is the one most likely to own the next wave of model training, agent improvement, and workflow-automation data assets.
Key diligence questions:
- Does the company capture real execution trajectories?
- Are those trajectories stored in a structured schema?
- Does the company have user authorization and compliance rights to retain the data?
- Can the product remain model-agnostic?
- Is there a plausible acquisition path by a platform or vertical operator?
- Has the company entered high-value workflows rather than demo infrastructure?
Layer 4: Workflow Ownership — The Main Venture Battlefield
Workflow Ownership is the most important part of the AI application layer for venture investors. The reason is straightforward: companies that own the workflow can own the execution data; companies that own execution data can move toward outcome-based pricing.
The SaaS era sold tools into businesses. The AI era allows companies to embed more deeply into business processes, and in some cases operate the business itself. Traditional SaaS improves efficiency, while the customer still owns the process, staff, data, and result. An AI company that only sells tools is often limited to subscriptions, seats, or usage pricing. A company that owns the customer relationship, execution process, and feedback loop can become a new kind of service operator, transforming labor-intensive sectors into AI-native operating companies.
Anthropic’s launch of an enterprise AI services firm with Blackstone, Hellman & Friedman, and Goldman Sachs is a signal of this direction. Frontier-model commercialization is moving from tokens and APIs toward organizational capability: embedding models, engineering teams, and industry know-how directly into enterprise workflows. This resembles Palantir-style forward-deployed engineering more than traditional software distribution.
The investment implication is significant. Model APIs will face margin and pricing pressure over time. Token-embedded workflows are stickier because they connect model capability to customer processes, permissions, delivery, data, and outcome measurement. Frontier companies may create alpha through vertical integration into workflow ownership, rather than model performance alone.
How to Invest in Workflow Ownership
This layer should be evaluated through control, measurement, and unit economics.
1. Customer relationship
If the AI company sits on top of the customer’s internal systems, the incumbent enterprise retains the relationship and much of the data. If the company delivers outcomes directly or becomes deeply embedded in a core process, it has a better chance of owning the workflow.
2. Full execution chain
Advice, drafting, summarization, and retrieval are partial value pools. Stronger companies close the loop from task intake to execution, exception handling, delivery, and postmortem feedback.
3. Measurable outcomes
Workflow Ownership works best where results can be quantified: customer-support resolution rate, insurance loss ratio, accounting delivery efficiency, legal-review accuracy, recruiting match success, sales conversion, or trading-execution quality.
4. Unit-economic improvement
AI transformation must show up in delivery cost, gross margin, cycle time, revenue per employee, retention, expansion, and pricing power.
Why Rollups Matter
Vertical agents may become one of the largest application opportunities of the AI era, but the best vehicle will not always be a traditional software startup. A powerful path is the AI-enabled rollup: acquire traditional service businesses, control customers and workflows, then rebuild delivery with AI.
The advantages are clear. Selling tools to an accounting firm, insurer, or BPO often leaves data rights with the client. Operating the business directly gives the AI company access to historical cases, employee workflows, customer feedback, outcome data, and margin structure. This enables pricing models that traditional SaaS struggles to support: per ticket resolved, per audit completed, per risk underwritten, or per claim processed. AI transformation then appears directly in EBITDA and margin expansion rather than only in software usage metrics.
There is also a moat that the market often underestimates: licenses and regulatory assets. If an AI-native company only sells a tool, it can be caught by traditional service firms or by model companies moving downstream. But if it also owns licenses, customer relationships, compliance capability, and the execution system, defensibility becomes much stronger. Corgi's acquisition of a full insurance-carrier license and Crete's control of multiple accounting-partner platforms are not just "software plus services." They are composite structures made of regulatory moat, workflow moat, and AI execution moat.
Crete Professionals Alliance, Corgi, and Crescendo + PartnerHero point to the same structural shift: AI is blurring the boundary between software companies and service companies.
There is also a lighter-weight route: instead of running the rollup directly, sell tools into PE-backed rollups. Many private-equity rollups already control large networks of operating locations and have a mandate for margin improvement. What they lack is AI workflow tooling that can be plugged in quickly. For startups, that can be a faster path than cold-starting one industry customer at a time. For investors, these companies are effectively serving the rollup ecosystem, not generic SaaS demand.
Which Workflow Ownership Opportunities Are Worth Watching
The strongest Workflow Ownership opportunities usually share six traits:
- High-frequency processes
- Labor intensity
- Measurable outcomes
- Willingness to pay for results
- Industry fragmentation
- Clear room for AI-led margin improvement
Attractive categories include:
| Category | Why It Matters |
| Customer support and BPO | Resolution rate, response time, and CSAT are measurable |
| Accounting and tax | Standardized processes, clear deliverables, fragmented market |
| Insurance and claims | Data-intensive, rules-intensive, outcome-linked |
| Legal and compliance | Document-heavy, process-complex, expensive to get wrong |
| Recruiting and sales | Trackable outcomes, though distribution matters heavily |
| Financial research and trading execution | Clear outcomes, high compliance and risk-control demands |
The caution zone includes vertical agents with only UI and prompt workflow, limited customer data, and weak outcome loops. They may look like applications, but they are often only short-term packaging around model capability. . AI tools sold purely for internal enterprise use, without access to execution data, can still become good SaaS businesses, but they do not necessarily own the workflow. Embedded AI that relies too heavily on model differentiation will also be squeezed as model companies ship native features.
Additionally, the rollup narrative requires discipline. Without M&A capability, financial integration, process redesign, AI implementation, and operating expertise, an AI rollup can become a traditional PE rollup with a layer of automation marketing on top.
The central underwriting question for this layer: does the company own the workflow, the outcome, and the data loop strongly enough to improve unit economics over time?
Layer 5: Agent Commerce + Identity — Crypto’s Strategic Opening
Agent Commerce + Identity is the most natural intersection between AI and crypto.
Agents need to discover services, call those services, verify identity, obtain permissions, pay for usage, settle transactions, and leave auditable records. Crypto and open-finance systems have spent the last decade building around related problems: digital identity, programmable payments, stablecoin settlement, wallet permissions, open protocols, and verifiable records.
However, most "AI tokens" and "agent tokens" are still narrative assets rather than real commerce infrastructure. The truly investable layer is infrastructure: payment rails, identity and reputation, service discovery, marketplaces, protocols, and developer standards.
The standards race is moving quickly. MCP has become one of the de facto standards for connecting agents to tools and external systems. A2A is pushing cross-framework communication between agents. x402 and Agentic.Market signal Coinbase’s ambition in agent commerce. Stripe, Visa, PayPal, Google, Cloudflare, Anthropic, and others will compete for payments, identity, and trust surfaces.
The highest near-term priority is KYA — Know Your Agent.
In financial services, payments, trading, and enterprise systems, non-human identity density is rising quickly. Agents, bots, service accounts, API clients, and automated workflows will soon outnumber human employees in many systems. Over the next 12 to 18 months, agent identity may become a basic requirement for agent commerce.
KYA must answer four questions:
- Whom does this agent represent?
- What is it authorized to do?
- Where are the permission boundaries?
- Who is accountable when something goes wrong?
Without identity and accountability, agent payments, marketplaces, machine-to-machine transactions, and automated trading all face a trust bottleneck.
Foresight’s stance: avoid rebuilding standards already being captured by large platforms. The better hunting ground is around still-unlocked implementation layers:
- Cross-chain, cross-fiat, and stablecoin settlement connectors
- Agent identity and KYA implementation
- Wallet permissions, session keys, and delegated payments
- Vertical agent-service marketplaces
- Trading-agent execution and risk-control harnesses
- Enterprise audit, permissioning, compliance, and risk management
- Cross-rail settlement between crypto, fiat, and stablecoins
For Bitget, this layer is strategically important. Exchanges and wallets naturally sit on the funding and execution paths of agent commerce. If agents trade for users, pay for services, subscribe to data, allocate into RWA, or settle cross-border transactions, competition will depend on which platforms become the default account, liquidity, and settlement layers that agents call.
Exchanges should therefore move beyond AI chatbots. The more strategic build areas are agent-native accounts, permission systems, execution traces, risk controls, KYA capabilities, and settlement infrastructure.
Foresight's Allocation View: Own Fewer Generic AI Bets, More Ownable Workflows
Across the five layers, Foresight suggests one unifying diligence formula:
Output per unit token = revenue uplift, cost savings, error reduction, or risk reduction divided by model and execution cost
The market previously focused on token call volume, API usage, user activity, and agent execution counts. The next phase will focus on what each dollar of model spend produces: revenue created, labor hours saved, errors prevented, conversion improved, or risk losses reduced.
This is especially important for Workflow Ownership. If a downstream AI product cannot present a clear outcome metric, it is vulnerable in enterprise budget reviews. The resilient companies will prove that AI spend converts into measurable business outcomes.
From an allocation perspective, Foresight's strategy can be summarized in five points.
1. Physical Substrate: Back load-bearing pillars
Power, cooling, packaging, networking, and leading neocloud players deserve sustained attention. Early hardware startups should be approached selectively.
2. Frontier Intelligence: Avoid general-model crowding
General models are already concentrated. The more interesting openings are Physical AI, World Models, embodied intelligence, robotics, and selected bio-frontier opportunities.
3. Harness: Prioritize closed loops and M&A optionality
Generic middleware is exposed to absorption by model platforms and vertical agents. The strongest Harness companies will either own execution trajectories inside high-value workflows or become strategic acquisition targets.
4. Workflow Ownership: Focus on control
The best opportunities own the customer, process, data, and outcome. AI-enabled rollups, services-as-software, and vertical operators may produce stronger moats than traditional SaaS.
5. Agent Commerce + Identity: Build strategic positioning for Foresight and Bitget
The investable themes are payments, identity, permissions, service discovery, cross-rail settlement, agent-native accounts, and execution infrastructure. Vague AI-token narratives should be filtered out.
If the strategy must be reduced to one sentence:
The next phase of AI investing is about backing companies that own the execution chain, the data loop, and the standards surface.
Conclusion: The AI Value Chain Is Being Reallocated
The first phase of AI belonged to models. The second phase belongs to companies that turn models into execution systems. The third phase will belong to companies that own workflows, data loops, and standards networks.
The “models vs. applications” binary cannot describe that reallocation with enough precision.
The better questions are:
- Who owns the physical bottlenecks?
- Who controls frontier intelligence?
- Who accumulates execution trajectories?
- Who owns the customer and workflow?
- Who defines the standards for agent commerce and identity?
Foresight’s five-layer framework separates those questions and links each layer to its own moat and diligence method.
The largest venture opportunities in the next AI cycle are likely to sit in companies that own high-value workflows. These companies may look less like traditional SaaS and more like PE rollups, BPOs, insurers, accounting firms, trading-execution platforms, or payment networks. Their advantage comes from entering real business operations, owning execution data, and compounding outcome data over time.
For Foresight and Bitget, this is both a financial-investment thesis and a strategic route. In the agent economy, trading, payments, identity, asset distribution, and execution networks will be redefined. The priority is to back the layers that can reshape financial entry points, execution chains, and capital formation.