From Tokens to Value: How Context Becomes the New AI Bottleneck (AI Supercycle’s Solow Paradox)

Key Insights

Enterprise AI bottleneck has moved to Context. Most large‑scale AI failures were once blamed on insufficient model intelligence; now the limiting factor is the lack of formalized decision‑making logic and process‑level knowledge within organizations.

Process redesign yields higher marginal returns than model upgrades. Adding AI to existing workflows yields modest gains, but reorganizing processes can increase output by an order of magnitude.

Agent trace compresses the Solow lag. Each AI Agent run automatically generates raw material for the process layer, allowing semi‑automatic accumulation of organizational knowledge and shortening deployment cycles.

Value stack inversion will happen, but slower than expected. Current AI value concentrates in the Semi layer; open‑source models pressure margins, pushing value upward, yet without true organizational digestion the Apps layer cannot take off.

1. AI Valuation Surge vs SaaS Valuation Low

In May 2026 the median EV/Revenue multiple for cloud software fell to 3.3× – a historic low, even lower than the 2022 rate‑hike panic. Meanwhile the four US giants (Google, Amazon, Microsoft, Meta) plan roughly $7.25 trillion in AI capex for 2026, and data‑center revenue for NVIDIA, Anthropic, and OpenAI remains in the hundreds of billions, with consumer AI WAU exceeding 1.2 billion but still not translating into measurable enterprise productivity.

“We’ve had this four times — Brexit, taper tantrum, inflation. This time it’s AI. Is this a buy‑the‑dip situation?” – Apoorv Agrawal

Ali Ghodsi replies that software is not dead, but differentiation will be severe. Companies that have not innovated in the past decade risk being displaced by AI‑native firms, while those with data, customers, and a willingness to restart innovation will survive.

2. Enterprise AI Bottleneck Lies in Context

Despite rapid advances in large‑model capabilities (e.g., Claude Fable 5, GPT‑5.5), MIT NANDA Lab’s 2025 report shows a 95 % failure rate for enterprise AI POCs, and Databricks reports no large‑scale agentic co‑workers among its 20 000 customers.

Ghodsi explains that models lack the internal organizational context—decision logic, workflow dependencies, tacit knowledge, and permission boundaries—that is never present in training data or API calls.

Gartner now proposes a two‑layer framework: a Knowledge Graph covering *what*/*who* (semantic layer) and a Context Graph covering *how*/*why* (process layer).

Semantic layer (What/Who): Structured data (warehouses, BI metrics) plus unstructured content (documents, emails, Slack) plus entity relationships. Databricks Lakehouse + Unity Catalog address the structured side; Glean Enterprise Graph handles unstructured content.

Process layer (How/Why): Captures explicit and implicit decision traces, including triggers, reference context, rules, and approval chains. Traditional Systems of Record never captured this. Two dimensions are decision logic and governance boundaries.

Most enterprises only implement the semantic layer, leaving the process layer empty, which explains why 95 % of POCs fail.

3. Process Redesign Beats Model Upgrades

Databricks’ connector team needed to integrate 800+ data‑source connectors. A traditional approach estimated nine months per connector.

First attempt – tool overlay, unchanged process: Adding a Copilot‑like tool reduced the timeline from nine months to 7.5 months.

Second attempt – first‑principles process redesign: Three changes were made:

Requirement gathering: From a quarter to one week, leveraging rapid code generation to rewrite specs cheaply.

Test environment: Outsourced parallel setup of Salesforce/Workday instances, eliminating a bottleneck.

Team structure: From 1:1 ownership to 7:7 cross‑coverage, removing single‑point‑of‑failure risk.

Result: seven connectors delivered per quarter—a 21× increase in unit output.

This mirrors Paul David’s 1990 observation that shifting from a group‑driven to a unit‑driven organization (electric motor vs. central shaft) unlocks productivity.

4. Organizational Redesign Has No Technical Shortcut

“You can see the computer age everywhere but not in the productivity statistics.” – Robert Solow, 1987

Historical parallels (electric motor, PC) show long lags between technology availability and productivity gains. The same lag appears for AI: the “Solow lag” will compress only the traceable, high‑frequency, rule‑based processes, not high‑level strategic decisions.

Three frictions impede organizational redesign:

Cognitive inertia: Management does not know what the new process should look like.

Sunk cost: Existing systems and teams built around legacy logic.

Institutional constraints: Compliance, unions, procurement, etc.

Technology alone cannot eliminate these.

5. Value Stack Inversion Is Slow

Jensen Huang’s five‑layer AI stack (Energy → Chips → Infrastructure → Models → Applications) shows current AI value concentrated in the Semi layer (~$300 B), with Infrastructure at $750 B and Apps at $600 B—an inversion compared to traditional software where value sits at the top.

Databricks’ product evolution (Spark → Lakehouse → Unity Catalog → Agent Bricks → Lakebase) illustrates a move from data storage/computation toward enabling AI to understand and act on organizational knowledge.

Open‑source model margins compress the Semi layer, forcing value upward, but without organizational capability the Apps layer cannot rise.

6. Three Layers of Friction from Feasibility to Value

Layer 1 – Model capability: Largely resolved after Claude Fable 5 and GPT‑5.5; frontier models can handle most enterprise tasks.

Layer 2 – Knowledge structuring: Decision logic remains scattered; companies like Databricks, Glean, and Palantir are building semantic and process layers via agent trace and manual ontology engineering.

Layer 3 – Organizational redesign: The connector case shows a 21× uplift only when CEOs drive change, teams execute, and outputs are highly standardized. Traditional firms face the three frictions above; agent trace can capture “how” but not “why”.

Progress across the three layers is asynchronous: model capability is near‑complete, knowledge structuring is advancing, but organizational redesign lags.

Signals that the third layer may be loosening include rapid ARR growth (>200 % YoY) for vertical AI applications that truly take over decisions, and mergers driven by process‑restructuring advantages.

Ultimately, AI will compress the Solow lag for traceable, high‑frequency processes, but the overall shift of value to the application layer depends on how quickly organizations can internalize AI‑generated context.