The Post-AGI Investment Thesis: Where Capital Flows After the Foundation Model Wars
The Post-AGI Investment Thesis: Where Capital Flows After the Foundation Model Wars
"In every infrastructure revolution, the railroads capture the headlines and the land speculators capture the returns. The question is always: what is the land?" — ThinkForge Research Brief, Q2 2026
00. Transmission Header
CLASSIFICATION : Tresslers Group Intelligence // ThinkForge Division
DOMAIN : Investment Intelligence / Capital Allocation / AI Market Structure
STATUS : Active Intelligence — Verified Telemetry
DATE : 2026.05.10
CAPITAL DATA : 2025 AI VC: $202–211B (49% of all global VC)
Foundation models: ~$80B raised (2025)
Hyperscaler capex: $350–410B combined (2025)
Agentic AI: $1.5B (2024) → $2.9B (2025)
ALERT LEVEL : High — Capital misallocation risk elevated; thesis differentiation critical
Every major technology infrastructure wave produces the same capital allocation error at scale: investors pour money into the most visible, most headline-generating layer of the stack — while the durable value accrues in the unglamorous infrastructure and application layers that no one is covering.
In the railroad era, it was the tracks. In the internet era, it was the protocols and browsers while fortunes were made in search and e-commerce. In the mobile era, it was handset manufacturers while the app stores and platform operators extracted the value.
In the AI era, the visible, headline-generating layer is foundation models. The durable value is somewhere else.
This dossier maps where.
01. The State of Capital: 2024–2025 Verified
The AI capital market has reached a scale that demands historical framing.
Verified 2025 investment data:
| Capital Category | 2024 | 2025 | YoY Change |
|---|---|---|---|
| Global AI VC funding (total) | $100–114B | $202–211B | ~+100% |
| AI share of all global VC | ~34% | ~49% | +15pp |
| Foundation model funding | ~$31B | ~$80B | +158% |
| AI infrastructure (IT + hosting) | $47.4B | $109.3B | +130% |
| Agentic AI disclosed funding | ~$1.5B | ~$2.9B | +93% |
| Hyperscaler combined capex | ~$200B | $350–410B | ~+80% |
The concentration problem: a small number of foundation model labs captured a disproportionate share of the total. OpenAI and Anthropic alone accounted for approximately 14% of all global venture investment in 2025. By some analyses, mega-rounds ($100M+) comprised 73% of total AI investment value in 2025 — meaning the headline numbers are heavily distorted by a handful of enormous financings.
Rendering diagram...
At these capital volumes — $202–211B in VC plus $350–410B in hyperscaler capex — the AI infrastructure build-out rivals the internet infrastructure investments of the late 1990s in scale. The question that followed the 1990s infrastructure boom is the same one now: who captures the return on this investment, and over what timeframe?
02. The Foundation Model Layer — Why It Is Not the Investment
The foundation model layer is where the majority of AI venture capital is concentrated, and it is not where the durable investment returns will accrue. This is a structural argument, not a qualitative judgment about model quality.
The commoditization trajectory:
In 2022, a frontier-capable large language model required proprietary training infrastructure, a team of hundreds of ML researchers, and compute budgets only accessible to a handful of organizations globally. The capability gap between frontier and open-source models was enormous.
By 2025, that gap has compressed materially:
- ▸Llama 3.1 405B (Meta, open-source): competitive with GPT-4 class models on many benchmarks, available for self-hosting at enterprise scale
- ▸Mistral and DeepSeek demonstrated that models within 10–15% of frontier performance can be trained at a small fraction of the cost
- ▸Inference costs dropped approximately 97% from 2022 to 2025 for equivalent capability levels — prices that would have been $30–50 per million tokens are now $0.15–3.00
The economic structure of a market with this cost trajectory is not one where the model producers capture durable margin. It is one where:
- ▸A small number of frontier models differentiate on capabilities at the leading edge (GPT-5, Claude Opus 4, Gemini Ultra 2)
- ▸The vast majority of production applications use commodity-priced models at the level below the frontier
- ▸Model providers compete primarily on price and context window size — not on defensible moats
The Microsoft/OpenAI analogy: Microsoft dominates enterprise software (Office, Azure, Teams) not because it builds the fastest CPUs. It builds on top of commoditized silicon while owning the application and distribution layer. The parallel in AI: the companies that own the application layer, the distribution, and the domain-specific deployments will capture returns proportional to their customer relationships and switching costs — not their model training compute.
03. The Five Layers of the AI Stack — Value Accrual Analysis
Rendering diagram...
Layer 1 — Silicon: Nvidia holds a dominant position in AI training compute. The H100 and H200 GPU families have been the primary compute substrate for frontier model training. However, custom ASIC development by hyperscalers (Google TPUs, AWS Trainium, Meta MTIA) is progressing rapidly, and each custom chip deployment reduces Nvidia's addressable market. The silicon layer has high current margin but faces long-term erosion from vertical integration.
Layer 2 — Compute Infrastructure: Data center buildout at $350–410B combined hyperscaler capex in 2025 represents a massive capital allocation to physical assets. These assets are durable — they do not depreciate as fast as software moats — but they also attract intense competition. The hyperscaler cloud layer will likely consolidate to three dominant providers (AWS, Azure, GCP) with commodity-pricing dynamics applying to inference by 2027.
Layer 3 — Foundation Models: See Section 02. High current valuations; commoditization trajectory structurally limits return on invested capital at scale.
Layer 4 — Orchestration and Infrastructure: This is the picks-and-shovels layer that most investors are underweighting. The tools required to build, deploy, monitor, and manage production AI agent systems — vector databases, evaluation frameworks, observability tools, orchestration frameworks — have low current valuations relative to their strategic importance. These tools also exhibit classic B2B SaaS characteristics: high switching costs once integrated into production workflows, expanding use cases as agent deployments scale, and pricing leverage tied to the value they enable rather than their own cost of production.
Layer 5 — Vertical Applications and Agent Deployments: This is the highest-return layer of the stack, and also the hardest to invest in at early stages because it requires deep domain expertise. The key insight from 2025 data: vertical AI agents — agents specialized for specific business outcomes in finance, healthcare, legal, procurement, etc. — grew from 48% to 72% of all agentic AI investment in one year. Investors are discovering that a general-purpose agent is an expensive commodity, but a specialist agent with proprietary domain data and validated performance is an extremely defensible business.
04. Where Tresslers Group Sees the Value: The Investment Thesis
The Tresslers Group investment thesis for the AI era is organized around three convictions:
Conviction 1: Vertical agent operators will be the professional services firms of the 2030s.
The professional services industry — consulting, research, legal, financial advisory — is structured around human expertise. The billable hour is the unit of value. The client relationship is the moat.
Vertical AI agent operators will replace large portions of this model with a structurally superior cost structure: agents that operate 24/7, produce consistent outputs at a small fraction of the cost, and become more accurate over time as their knowledge substrate deepens. The organizations that own these deployments — that have the proprietary domain data, the validated performance track records, and the enterprise relationships — will capture margin that currently flows to professional services headcount.
Conviction 2: The data substrate is the moat.
A generic AI agent has no moat. A specialist agent operating against a proprietary, continuously-updated, domain-specific knowledge base — clinical literature, trade intelligence, regulatory filings, patent databases — has a moat that compounds. Each query enriches the knowledge base. Each deployed hour increases the gap between the specialist and the generic alternative.
The Intelligence Library being built at Tresslers Group is not a publishing exercise. It is the data substrate construction that makes the agent fleet categorically better than alternatives. This is the "land" in the railroad analogy.
Conviction 3: Infrastructure for the agentic economy will be underpriced until it isn't.
The agentic AI infrastructure layer — orchestration frameworks, MCP tool registries, agent wallet infrastructure (x402), evaluation and observability platforms — is currently receiving approximately 1.4% of total AI VC ($2.9B in 2025). When agent deployment becomes mainstream enterprise practice (estimated 2027–2028 by most analysts), the infrastructure layer becomes critical path — and reprices accordingly.
The investment window for this layer is 18–36 months. After that, it looks like enterprise SaaS in 2015: everyone knows it's important and valuations reflect that.
05. The Capital Misallocation Map
Understanding where capital is currently misallocated allows identification of the opportunities that remain underpriced:
| Category | Current Capital Allocation | Structural Value | Misallocation Signal |
|---|---|---|---|
| Foundation model training | Very High ($80B VC + hyperscaler training CapEx) | Medium (commoditizing) | Overvalued relative to durable moat |
| Data center/compute | Very High ($109B IT infrastructure VC + $350–410B capex) | High (physical assets) | Appropriately priced — physical scarcity |
| GPU silicon | High (Nvidia market cap ~$3T peak) | Medium-High | Partially overvalued vs. custom ASIC threat |
| Orchestration / MLOps | Medium ($5–15B estimated) | High | Undervalued — picks-and-shovels |
| Vertical agent operators | Low ($2.9B total agentic, rapidly growing) | Very High | Significantly undervalued — early innings |
| Domain knowledge substrates | Very Low (not tracked as category) | Very High | Most undervalued category in the stack |
| Agent payment infrastructure (x402) | Minimal (nascent ecosystem) | High | Pre-market — first-mover window |
The highest durable value / lowest current capital allocation ratio is in domain knowledge substrates — proprietary, structured, continuously-updated knowledge bases that make specialist agent deployments categorically better than generic alternatives. This is not a category that appears in VC tracking databases. It is being built inside organizations that understand the structural dynamic before the market prices it.
06. The Agentic AI Funding Shift — Signal Analysis
The shift in agentic AI investment from horizontal to vertical agents (48% → 72% vertical in 2025) is one of the clearest market signals of where intelligent capital is moving.
Rendering diagram...
What this signal means: the market has rapidly learned that general-purpose agents are commodities. A "research agent" that uses OpenAI's API, browses the web, and produces reports is reproducible in a weekend. A specialist agent for pharmaceutical regulatory submissions that knows FDA 21 CFR Part 11, has ingested 10 years of approved NDA submissions, and produces compliance-validated documentation — that is a business with real switching costs.
The capital is following the moat. Tresslers Group's ThinkForge and Zoirah deployments are explicitly in the vertical agent category — domain-specific intelligence that compounds in value with each deployed hour.
07. Sector-by-Sector Capital Flow Analysis
Where AI investment is flowing by sector in 2025:
Healthcare and Life Sciences: The combination of AlphaFold 3 (May 2024), FDA's record 295 AI/ML device authorizations in 2025, and structural drug discovery timeline compression (25–70%) has made healthcare AI one of the highest-conviction investment categories. Capital is flowing to diagnostic AI, drug discovery platforms, and clinical operations automation. The TAM is enormous — global healthcare represents ~$10T annually.
Financial Services: AI in finance has been the fastest-moving vertical — hedge funds and trading desks deployed AI systems first, followed by retail banking, insurance underwriting, and financial planning. The market is bifurcating: generic fintech AI is becoming commoditized, while specialized financial intelligence agents with proprietary data feeds are building defensible positions.
Defense and Intelligence: Government and defense AI spending is largely invisible in public VC data but is substantial. The US Defense Department's AI spending exceeded $1.5B in FY2025, and allied nations are increasing allocations. This sector has the highest regulatory barriers and the most defensible moats.
Enterprise Automation: The displacement of SaaS by agent fleets (mapped in the Tresslers Group Agentic Supply Chain dossier) is creating investment opportunities in the companies providing orchestration infrastructure and domain-specific automation. The Gartner forecast of agentic SCM software growing from <$2B to $53B by 2030 is one data point in a broader pattern.
08. The Risks — Honest Assessment
The valuation risk: foundation model companies are carrying valuations that imply market structures that may not materialize. If open-source models continue to compress the capability gap — as Llama 3 and DeepSeek demonstrated in 2024–2025 — revenue models dependent on proprietary model access face structural pressure.
The compute cost risk: the hyperscaler capex build-out ($350–410B in 2025) is predicated on sustained AI demand growth. If AI application adoption is slower than projected, the industry faces an overcapacity condition analogous to the fiber optic build-out of the late 1990s.
The regulatory risk: the EU AI Act is in phased implementation. Additional regulatory frameworks in the US, UK, and Asia-Pacific are being developed. High-risk AI applications face compliance requirements that add cost and extend deployment timelines, potentially compressing the margin available in specific verticals.
The concentration risk: 73% of AI investment value in 2025 was in mega-rounds. This concentration means that reported market totals are driven by a handful of enormous financings — the underlying ecosystem of smaller, earlier-stage companies is less capital-dense than the totals suggest.
09. The Tresslers Group Thesis
The foundation model wars will produce winners. The returns will be captured elsewhere.
This is not a prediction that foundation model companies will fail. Some will succeed enormously. OpenAI, Anthropic, and Google DeepMind are building capabilities that will define the infrastructure of the next decade.
The investment thesis is simpler: in every infrastructure revolution, the infrastructure providers are necessary but not sufficient for durable value capture. The durable value accrues to those who own the customer relationship, the domain expertise, and the proprietary data that makes their application of the infrastructure categorically better than generic alternatives.
The railroad companies that survived did so because they owned routes with structural competitive advantages — geography, regulatory protection, network effects. In AI, the equivalent is domain-specific data substrates, validated agent performance track records, and enterprise relationships that create switching costs.
Tresslers Group is building in the vertical agent operator category — across ThinkForge (research and intelligence), Zoirah (healthcare), and Tressler's Trading (commerce) — because this is where the thesis plays out with the clearest line from capability to durable competitive advantage.
Capital follows the signal. The signal says: vertical, domain-specific, data-substrated.
References & Source Intelligence
- ▸Crunchbase. (2025). Global AI Funding: $202–211B in 2025, 49% of All Global VC.
- ▸Venture Capital Journal. (2025). Foundation Model Funding: $80B in 2025, 2× 2024.
- ▸OECD. (2025). AI Infrastructure Investment: IT and Hosting Firms Attract $109.3B in 2025.
- ▸NewMarketPitch. (2025). Agentic AI Funding: $1.5B (2024) to $2.9B (2025); Vertical Agents Rise to 72%.
- ▸AWS / Microsoft / Google / Meta. (2025). Annual Reports and Earnings Calls: Combined CapEx $350–410B.
- ▸Tresslers Group Intelligence. (2026). The Agentic Supply Chain. [tresslersgroup.com/insights/agentic-supply-chain-2026]
- ▸Tresslers Group Intelligence. (2026). Agent-to-Agent Commerce: The x402 Economy. [tresslersgroup.com/insights/agent-commerce-x402-economy]
- ▸Tresslers Group Intelligence. (2026). AI Diagnostics at Scale. [tresslersgroup.com/insights/ai-diagnostics-clinical-intelligence-2026]
- ▸Tresslers Group Intelligence. (2026). Supply Chain Sovereignty. [tresslersgroup.com/insights/supply-chain-sovereignty-2026]
Tresslers Group Intelligence — ThinkForge Division Driven by Innovation. Defined by Impact. Thesis-Driven Capital Intelligence. © 2026 Tresslers Group. Transmission Complete.