Report IV · Published March 27, 2026 · Updated July 9, 2026
Structured Test-Time
Scaling
Why structure, verification, and recursion matter more than just thinking longer
April 2026 Update
- OpenAI's GPT-5.4-Cyber and Anthropic's Claude Opus 4.7 both lean harder into verified, long-running agent workflows.
- On Apr. 16, arXiv's “Scaling Test-Time Compute for Agentic Coding” introduced Recursive Tournament Voting and Parallel-Distill-Refine for long-horizon coding agents.
- AWS Trainium's Apr. 15 speculative-decoding guide reports up to 3x faster token generation on decode-heavy workloads when the draft model can predict well.
Current Thesis Audit · Updated July 9, 2026
For structured test-time scaling, the current pass favors verifiable reasoning, long-horizon coding agents, security evaluation, and infrastructure/software names that benefit when inference shifts from chat to agentic workflows. The active ranking pass cross-checks the original thesis against current market prices, market cap, YTD rerating, AI 2027 compute/security assumptions, hyperscaler capex pressure, power constraints, HBM and advanced-packaging evidence, and physical-AI adoption signals.
0
Companies
Global universe
0
Scenarios
Lattice combos
0
RLM Methods
Novel proposals
0
Verified Claims
Evidence-backed
Chapter 1
The Thesis
The research yields four bottom lines. Each is backed by 15 individually scored claims drawn from recent arXiv papers, coding-agent results, and infrastructure data.
01
Directionally, the thesis is strong: structure, context isolation, and verification matter more than merely asking a model to think longer.
02
The overreach is theoretical certainty: the exponent claim is best read as a heuristic rather than a proven universal law.
03
Durable rents sit in bottlenecks and trust layers. New arXiv evidence shifts value toward recursive training data flywheels, state management, synthetic environments, and verifier infrastructure.
04
The next frontier looks less like manually prompt-engineered multi-agent systems and more like models that natively learn when to recurse, what to remember, and what to verify.
15 Claims: Interactive Verdict Table
Structured test-time scaling is a more promising frontier than simply lengthening a single chain-of-thought.
SupportedConfidence: High
▼
Naive longer-is-better CoT scales poorly and can reduce accuracy.
SupportedConfidence: High
▼
Topology compression (hierarchy / recursion / sub-agents) is the key reason structured systems avoid linear long-horizon collapse.
Conditionally supportedConfidence: Medium
▼
Scope isolation lowers error rates by reducing context noise and local task difficulty.
SupportedConfidence: High
▼
Verification is the decisive third layer that turns residual errors into bounded tails.
Supported in verifier-rich domainsConfidence: High
▼
Dynamic communication routing alone is not equivalent to true span compression.
SupportedConfidence: Medium
▼
The unified law reduces effective failure growth from Theta(W) to O(log W).
Overstated / not yet establishedConfidence: High
▼
Coding agents are the cleanest real-world proof point for the thesis.
SupportedConfidence: High
▼
Larger context windows alone will not remove the need for scope isolation.
Conditionally supportedConfidence: Medium
▼
If AI 2027-style acceleration happens, power, packaging, cooling, and export controls become the real bottlenecks.
SupportedConfidence: High
▼
AI 2027 should be read as a forecast of what will happen.
Not supportedConfidence: High
▼
The article is a finished paper with completed experiments.
FalseConfidence: High
▼
The frontier is shifting from scaffold-only agent systems toward models post-trained to manage recursion and state internally.
SupportedConfidence: Medium-High
▼
Reasoning can improve without additional human labels via self-play, debate, or automated trajectory feedback.
SupportedConfidence: High
▼
The strongest medium-term moat may be better recursive controllers and verifier-rich synthetic training flywheels, not just prompt-engineered orchestration wrappers.
Conditionally supportedConfidence: Medium
▼
Chapter 2
Five Methods That Could Change Everything
Five novel RLM methods emerge from the literature. Each proposes a different route for structured test-time scaling to compound.
Verifier-First Recursive Compilation (VFRC)
Novelty: Medium-High▼
Before solving, the controller recursively compiles the task into contracts, tests, invariants, simulators, and typed outputs. It only spends large solve-budget once the task is maximally machine-checkable.
Counterfactual Call-Graph Credit Assignment (CCGA)
Novelty: High▼
Sample many decomposition trees for the same prompt, run them to verified completion, and assign regret/credit to split choices, stop decisions, memory writes, and model/tool routing. Train the recursive controller off-policy on that signal.
Recursive Compute Markets (RCM)
Novelty: Medium▼
Each branch predicts its verifier-adjusted value of additional compute (tokens, depth, model size, tool calls) and bids for budget. The parent allocates budget across branches and learns from realized payoffs.
Auto-Discovered Recursive Interfaces (ADRI)
Novelty: Medium-High▼
Train branches to invent compact typed schemas, DSL fragments, or latent APIs for recurring subproblems, constrained by round-trip reconstruction and verifier success. Parent-child communication shifts from verbose natural language to discovered interfaces.
Recursive Self-Play Environment Mutation (RSEM)
Novelty: Medium-High▼
The system generates synthetic tasks/environments that specifically require deeper recursion, memory hygiene, and stronger verification; it keeps only automatically scored tasks and ratchets up depth/branching as the model improves.
Chapter 3
The Strategic Chessboard
Twelve strategic games define the competitive landscape. Each pits different actors against each other with distinct equilibria and investment implications.
US vs China
▼
Each side wants leadership in chips, compute, and frontier models while fearing relative loss from restraint.
Frontier lab vs frontier lab
▼
Being second can be economically or strategically costly, so labs overspend on compute, talent, and data-center access.
Labs vs regulators
▼
Labs push speed; regulators threaten delay or restrictions after incidents.
Hyperscalers vs utilities / grids
▼
Cloud builders need power quickly; utilities need time, permits, transformers, and rate recovery.
Chipmakers vs hyperscalers
▼
Scarce HBM, advanced packaging, and leading-edge wafer capacity give suppliers temporary pricing power.
Closed-model platforms vs open ecosystems
▼
Closed platforms monetize reliability and integrated tooling; open ecosystems drive diffusion and price pressure.
Enterprises vs AI vendors
▼
Buyers delay large commitments until reliability, governance, and ROI are visible.
Data-center developers vs local communities
▼
Developers want scale and speed; communities worry about power, water, land, and rates.
Generator model vs verifier / critic
▼
The system wants speed and creativity from generators but must constrain silent error propagation.
Allies vs US export-control regime
▼
Allies want access to US tech but also local industrial upside and strategic autonomy.
Attackers vs defenders in AI-native software
▼
Autonomous tooling raises both attack surface and defensive automation potential.
Capital markets vs AI infrastructure builders
▼
Cheap capital chases AI capacity; project returns depend on utilization and power access.
Chapter 4
2026 – 2036: A Decade of Disruption
A year-by-year view of how the landscape evolves over the next decade. Scroll through the future.
2026
Structured inference proves itself first in coding, math, and research copilots; the first natively recursive and state-managing post-trained models appear, showing that some scaffold logic can move into the weights.
Indicators
RLM releases, state-aware memory training, verification-heavy research agents, more sub-agent tooling, and model-managed memory in major platforms
Risk
Gains stay narrow; recursive/stateful training proves brittle or too expensive outside verifier-rich domains
Beneficiaries
Evaluation/verification, coding agents, state management, synthetic trajectory tooling, advanced packaging, grid equipment
2027
Inference-time compute budgets rise, but the best labs increasingly train the controller itself: recursion depth, memory writes, and verification policies become learned rather than manually scripted.
Indicators
More papers on self-play/debate post-training, distillation of agent traces, adaptive compute allocation, model-managed memory, and benchmark wins from native recursive/stateful models
Risk
Off-policy training instability, weak verifiers, or excessive inference cost keep manual orchestration competitive
Beneficiaries
Verification/test infrastructure, synthetic environment generation, telemetry-rich orchestration, state stores, AI-native security
2028
The bottleneck shifts from model quality alone to infrastructure quality: power, cooling, HBM, advanced packaging, and networking.
Indicators
Data-center power queues lengthen; HBM and packaging lead times remain strategic; liquid-cooling capex accelerates
Risk
Overbuild in some regions or a chip-cycle downturn compresses pricing
Beneficiaries
Power developers, transformers, cooling, optical interconnect, packaging/metrology
2029
Sovereign and regional AI stacks emerge as export controls, trusted supply chains, and data-locality rules shape deployment.
Indicators
More national AI infrastructure programs, sovereign cloud/compute partnerships, export-control revisions
Risk
Geopolitical de-escalation or technological substitution reduces fragmentation
Beneficiaries
Non-US sovereign-stack partners, European/Japanese/Korean/Taiwanese suppliers, compliance software
2030
For many enterprise workloads, inference operating cost becomes as strategically important as training cost.
Indicators
Utilities cite data centers as a major demand driver; enterprises optimize for efficient verified agents rather than raw model size
Risk
Cheaper compute or highly efficient architectures blunt the infra bottleneck
Beneficiaries
Efficient inference silicon, power-management, context/agent middleware, enterprise observability
2031
Winning products behave like 'AI operating systems': orchestration, memory, policy, verification, and tool execution bundled together.
Indicators
Mature agent platforms offer routing, memory, subagents, evaluation, audit trails, and policy controls out-of-the-box
Risk
Platform consolidation squeezes independents; hyperscalers subsume the middleware layer
Beneficiaries
Open but enterprise-ready middleware, trust/security, workflow integration vendors
2032
High-stakes industries require auditable verification and formal assurance layers before large-scale autonomy is allowed.
Indicators
Procurement standards for traceability, test evidence, model risk management, and human override
Risk
Regulation lags or remains fragmented across jurisdictions
Beneficiaries
Verification, compliance, cyber, model observability, industrial QA
2033
Physical-world deployment expands: robotics, industrial automation, and edge inference benefit from structured reasoning plus localized verification.
Indicators
More on-device/edge AI accelerators, industrial MLOps, robotics with tighter task-specific feedback loops
Risk
Robotics adoption remains slower than software due to safety and labor/process friction
Beneficiaries
Robotics, edge AI silicon, sensors, industrial AI ops, safety infrastructure
2034
Model-layer margins compress if open models and commodity orchestration spread; durable rents move down-stack and into regulated trust layers.
Indicators
Open-weight adoption, price competition, enterprise preference for audited integrations over premium raw tokens
Risk
A few frontier labs preserve moat via superior verifier-rich systems or exclusive data/compute
Beneficiaries
Infrastructure, security, compliance, data rights, sovereign/local providers
2035
AI supply chains regionalize further around allied blocs for chips, energy, data, and security assurance.
Indicators
Regional fabs, packaging sites, power-heavy AI campus deals, industrial policy tied to strategic tech
Risk
Trade normalization or new architectures reduce dependency on current chokepoints
Beneficiaries
Allied non-US suppliers in semis, packaging, grid, and secure infrastructure
2036
The long-run winners are not just model creators but operators of reliable inference systems: power, packaging, cooling, orchestration, and verification.
Indicators
Enterprise spend concentrates on uptime, auditability, efficiency, and sovereign control rather than raw benchmark wins
Risk
A discontinuous model breakthrough re-centralizes value at the frontier model layer
Beneficiaries
Companies with bottleneck control, verification trust, or sovereign deployment advantages
Chapter 5
Scenarios: 10 Worlds, 16 Combinations
Ten base scenarios map to 16 lattice combinations across four axes: geopolitics, resource availability, architecture paradigm, and market structure.
10 Base Scenarios
Managed race, ample power
MediumUS-China competition stays intense but below crisis level; utilities and independent power keep up enough to avoid severe rationing.
Managed race, power bottleneck
HighDemand for AI campuses outruns grid upgrades, transformers, and cooling capacity.
Export-control ratchet
HighSemiconductor and model-stack controls tighten further; trusted-bloc procurement becomes strategic.
Packaging / HBM squeeze
HighCompute exists on paper but advanced packaging and HBM remain the scarce layer.
Cooling becomes the gating factor
Medium-HighThermal density rises faster than conventional air-cooling can handle.
Verification-first enterprise adoption
HighEnterprises adopt AI fastest where outputs are testable, auditable, and attributable.
Open-model diffusion / model margin compression
MediumModel quality diffuses broadly and pricing compresses at the raw model layer.
Sovereign stack build-out
Medium-HighEurope, Middle East, India, and parts of Asia accelerate local compute, cloud, and data-governance ecosystems.
Security incident / trust shock
MediumA major cyber, model poisoning, or agent misuse event drives a trust reset.
Overbuild / utilization bust
MediumCapital floods into capacity before steady demand materializes, hurting undifferentiated operators.
16-Cell Scenario Lattice
Geopolitics
Resource
Architecture
Market
S01Medium-High
Coordinated competition · Resource slack · Structured + verified · Concentrated / closed
Winning: Agent middleware, verification, workflow software, selected semis, security
Lean toward software infrastructure, observability, and enterprise automation.
S02Medium-High
Coordinated competition · Resource slack · Structured + verified · Diffuse / open
Winning: Agent middleware, verification, workflow software, selected semis, security
Lean toward software infrastructure, observability, and enterprise automation.
S03Medium
Coordinated competition · Resource slack · Brute-force / monolithic · Concentrated / closed
Winning: Agent middleware, verification, workflow software, selected semis, security
Favor differentiated platforms and cost leaders; avoid me-too wrappers.
S04Medium-High
Coordinated competition · Resource slack · Brute-force / monolithic · Diffuse / open
Winning: Agent middleware, verification, workflow software, selected semis, security
Favor differentiated platforms and cost leaders; avoid me-too wrappers.
S05Medium
Coordinated competition · Resource bottleneck · Structured + verified · Concentrated / closed
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Best barbell: bottleneck hardware + verification/trust software.
S06Medium-High
Coordinated competition · Resource bottleneck · Structured + verified · Diffuse / open
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Best barbell: bottleneck hardware + verification/trust software.
S07Medium
Coordinated competition · Resource bottleneck · Brute-force / monolithic · Concentrated / closed
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Favor hard bottlenecks over app-layer exposure.
S08Medium
Coordinated competition · Resource bottleneck · Brute-force / monolithic · Diffuse / open
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Favor hard bottlenecks over app-layer exposure.
S09Medium
Hard race · Resource slack · Structured + verified · Concentrated / closed
Winning: Agent middleware, verification, workflow software, selected semis, security
Lean toward software infrastructure, observability, and enterprise automation.
S10Medium-High
Hard race · Resource slack · Structured + verified · Diffuse / open
Winning: Agent middleware, verification, workflow software, selected semis, security
Lean toward software infrastructure, observability, and enterprise automation.
S11Medium
Hard race · Resource slack · Brute-force / monolithic · Concentrated / closed
Winning: Agent middleware, verification, workflow software, selected semis, security
Favor differentiated platforms and cost leaders; avoid me-too wrappers.
S12Medium
Hard race · Resource slack · Brute-force / monolithic · Diffuse / open
Winning: Agent middleware, verification, workflow software, selected semis, security
Favor differentiated platforms and cost leaders; avoid me-too wrappers.
S13Medium
Hard race · Resource bottleneck · Structured + verified · Concentrated / closed
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Best barbell: bottleneck hardware + verification/trust software.
S14Medium
Hard race · Resource bottleneck · Structured + verified · Diffuse / open
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Best barbell: bottleneck hardware + verification/trust software.
S15Low-Medium
Hard race · Resource bottleneck · Brute-force / monolithic · Concentrated / closed
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Favor hard bottlenecks over app-layer exposure.
S16Medium
Hard race · Resource bottleneck · Brute-force / monolithic · Diffuse / open
Winning: Power, cooling, packaging, grid equipment, efficient silicon, security
Favor hard bottlenecks over app-layer exposure.
Chapter 6
100 Companies to Watch
Split evenly: 50 public names and 50 private ones. Filter by category, sort by any column, and expand rows for risk detail.
Bucket Distribution
| # ▲ | Company | Ticker | Country | Exchange | Bucket | Residual | Cases | Price / Cap | Why It Fits |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Voltamp Transformers | VOLTAMP.NS | India | NYSE/NASDAQ | Electricity | 97.5 | INR 9,580 $1.0B +23.0% YTD | Transformer bottleneck exposure | |
| 2 | TOWA Corporation | 6315.T | Japan | TSE | Semiconductors | 90.5 | JPY 2,993 $1.6B +30.9% YTD | Packaging/molding equipment for semiconductor backend | |
| 3 | Tekscend Photomask | 429A.T | Japan | TSE | Semiconductors | 82.5 | JPY 4,300 $2.9B +39.6% YTD | Photomask/reticle infrastructure for advanced chip production | |
| 4 | Cohu | COHU | United States | NYSE/NASDAQ | Semiconductors | 76.4 | $52.12 $2.4B +111.8% YTD | Test and inspection gear; verification layer for chip manufacturing | |
| 5 | Scott Technology | SCT.NZ | New Zealand | NYSE/NASDAQ | Robotics | 69.1 | NZD 2.76 $128M -4.5% YTD | Industrial automation and robotics systems | |
| 6 | Valens Semiconductor | VLN | Israel | NYSE/NASDAQ | Semiconductors | 69.0 | $2.04 $258M +30.8% YTD | High-speed connectivity silicon | |
| 7 | Inox Wind | INOXWIND.NS | India | NYSE/NASDAQ | Electricity | 62.9 | INR 83.19 $1.6B -32.3% YTD | Wind generation equipment and power capacity build-out | |
| 8 | Voltalia | VLTSA.PA | France | Euronext | Electricity | 62.3 | EUR 7.15 $1.1B -12.4% YTD | Renewable power developer | |
| 9 | Genesis Energy | GNE.NZ | New Zealand | NYSE/NASDAQ | Electricity | 61.5 | NZD 2.58 $1.7B +7.5% YTD | Utility / power demand exposure | |
| 10 | Romande Energie | REHN.SW | Switzerland | SIX | Electricity | 61.5 | CHF 49.3 $1.6B +14.1% YTD | Utility exposure to electricity demand growth | |
| 11 | Maytronics | MTRN.TA | Israel | NYSE/NASDAQ | Robotics | 58.3 | n/a $83M n/a YTD | Robotic systems / consumer automation | |
| 12 | u-blox | UBXN.SW | Switzerland | SIX | Semiconductors | 57.8 | n/a $1.3B n/a YTD | Positioning and wireless modules for edge/robotic deployment | |
| 13 | PDF Solutions | PDFS | United States | NYSE/NASDAQ | Semiconductors | 57.1 | $52.42 $2.2B +77.4% YTD | Yield analytics / process data software for fabs | |
| 14 | SkyWater Technology | SKYT | United States | NYSE/NASDAQ | Semiconductors | 56.8 | $33.34 $1.7B +48.6% YTD | US specialty foundry leverage in sovereign AI supply chains | |
| 15 | Yubico | YUBICO.ST | Sweden | NYSE/NASDAQ | It Security | 55.9 | SEK 56.8 $533M -22.2% YTD | Identity/authentication for agentic workflows | |
| 16 | Serve Robotics | SERV | United States | NYSE/NASDAQ | Robotics | 53.7 | $5.89 $514M -50.2% YTD | Autonomous delivery optionality | |
| 17 | O.Y. Nofar Energy | NOFR.TA | Israel | NYSE/NASDAQ | Electricity | 53.2 | n/a $2.6B n/a YTD | Power build-out and renewable generation | |
| 18 | VIGO Photonics | VGO.WA | Poland | NYSE/NASDAQ | Scientific Instruments | 53.0 | PLN 560 $126M +18.6% YTD | Photonics and infrared detector leverage | |
| 19 | Kraken Robotics | PNG.V | Canada | NYSE/NASDAQ | Robotics | 52.6 | CAD 6.05 $1.3B -10.9% YTD | Autonomous sensing and robotics systems | |
| 20 | Nachi-Fujikoshi | 6474.T | Japan | TSE | Robotics | 51.8 | JPY 6,040 $879M +34.2% YTD | Industrial automation / robotics components | |
| 21 | X-FAB | XFAB.PA | Belgium | Euronext | Semiconductors | 49.0 | EUR 7.97 $1.2B +51.0% YTD | Specialty foundry for mixed-signal/industrial/auto compute edge | |
| 22 | Eos Energy Enterprises | EOSE | United States | NYSE/NASDAQ | Electricity | 48.9 | $4.51 $1.6B -65.3% YTD | Long-duration storage optionality | |
| 23 | Sensirion Holding | SENS.SW | Switzerland | SIX | Scientific Instruments | 47.5 | CHF 78.1 $1.6B +26.0% YTD | Sensors for thermal/industrial control | |
| 24 | NCC Group plc | NCC.L | United Kingdom | LSE | It Security | 47.0 | GBp 138 $519M +2.7% YTD | Cyber assurance and security testing | |
| 25 | Ekinops | EKI.PA | France | Euronext | Networking | 46.1 | EUR 2.42 $82M +25.8% YTD | Optical transport/access infrastructure | |
| 26 | Arteris | AIP | United States | NYSE/NASDAQ | Semiconductors | 45.4 | $34.46 $1.6B +121.2% YTD | On-chip interconnect IP; complexity scaling inside AI silicon | |
| 27 | Allot | ALLT | Israel | NYSE/NASDAQ | It Security | 43.3 | $8.54 $343M -11.6% YTD | Network intelligence / traffic control | |
| 28 | Aviat Networks | AVNW | United States | NYSE/NASDAQ | Networking | 43.3 | $20.66 $273M -4.3% YTD | Microwave backhaul / resilient enterprise network links | |
| 29 | F-Secure | FSECURE.HE | Finland | NYSE/NASDAQ | It Security | 42.5 | EUR 1.96 $411M +1.4% YTD | Endpoint and consumer/business security | |
| 30 | LumenRadio | LUMEN.ST | Sweden | NYSE/NASDAQ | Networking | 42.5 | SEK 61.8 $79M +8.4% YTD | Industrial wireless connectivity / edge control | |
| 31 | Himax Technologies | HIMX | Taiwan | NYSE/NASDAQ | Semiconductors | 42.2 | $14.9 $2.6B +74.7% YTD | Edge AI vision/display chips and interface silicon | |
| 32 | ChipMOS Technologies | IMOS | Taiwan | NYSE/NASDAQ | Semiconductors | 41.1 | $68.07 $2.3B +122.9% YTD | Backend packaging and test exposure | |
| 33 | Alpha & Omega Semiconductor | AOSL | United States | NYSE/NASDAQ | Semiconductors | 40.1 | $37.37 $1.3B +80.9% YTD | Power semiconductors and power management | |
| 34 | secunet | YSN.F | Germany | FSE | It Security | 39.8 | EUR 170.8 $1.3B -5.5% YTD | Sovereign cybersecurity / regulated deployment | |
| 35 | Ceragon Networks | CRNT | Israel | NYSE/NASDAQ | Networking | 39.5 | $2.32 $256M +5.9% YTD | Wireless backhaul and network densification | |
| 36 | JEOL Ltd. | 6951.T | Japan | TSE | Scientific Instruments | 39.5 | JPY 7,309 $2.3B +45.2% YTD | Scientific instrumentation / metrology and analysis | |
| 37 | Tejas Networks | TEJASNET.NS | India | NYSE/NASDAQ | Networking | 39.0 | INR 540.4 $1.0B +19.5% YTD | Telecom/networking equipment in sovereign build-outs | |
| 38 | Babcock & Wilcox | BW | United States | NYSE/NASDAQ | Electricity | 36.8 | $10.77 $2.4B +69.6% YTD | Power systems / thermal infrastructure | |
| 39 | Jenoptik | JEN.F | Germany | FSE | Scientific Instruments | 30.5 | EUR 39.18 $2.9B +96.1% YTD | Optical systems, metrology, photonics | |
| 40 | Siltronic | WAF.F | Germany | FSE | Semiconductors | 30.3 | EUR 82.45 $3.2B +64.9% YTD | Wafer substrate supply; upstream semiconductor bottleneck | |
| 41 | Wolfspeed | WOLF | United States | NYSE/NASDAQ | Semiconductors | 26.9 | $35.86 $1.9B +89.4% YTD | Power semiconductors for AI power-density growth; very high risk | |
| 42 | SÜSS MicroTec | SMHN.DE | Germany | XETRA | Semiconductors | 26.8 | EUR 81.6 $1.9B +99.8% YTD | Advanced packaging and lithography process tools | |
| 43 | A10 Networks | ATEN | United States | NYSE/NASDAQ | It Security | 23.1 | $37.09 $2.6B +113.8% YTD | Network security and application delivery for AI-heavy traffic | |
| 44 | Aehr Test Systems | AEHR | United States | NYSE/NASDAQ | Semiconductors | 13.0 | $67.89 $2.1B +206.4% YTD | Burn-in / reliability testing for power and AI-adjacent chips | |
| 45 | Formosa Sumco Technology | 3532.TW | Taiwan | TWSE | Semiconductors | 0.0 | TWD 423.5 $4.6B +337.5% YTD | Silicon wafer supply for AI-related semiconductor capacity | |
| 46 | Sterlite Technologies | STLTECH.NS | India | NYSE/NASDAQ | Networking | 0.0 | INR 514 $3.2B +401.5% YTD | Fiber and optical transport for AI data movement | |
| 47 | Kalray | ALKAL.PA | France | Euronext | Semiconductors | 0.0 | EUR 8.32 $167M +577.5% YTD | Data-processing / DPU-style optionality | |
| 48 | IQE plc | IQE.L | United Kingdom | LSE | Semiconductors | 0.0 | GBp 38.99 $801M +657.1% YTD | Epitaxy materials stack; upstream bottleneck exposure |
Chapter 7
The Corrected Conclusion
The bottom lines hold. But the arXiv evidence demands a correction in where durable value sits.
Directionally, the thesis is strong: structure, context isolation, and verification matter more than merely asking a model to think longer.
The overreach is theoretical certainty: the exponent claim is best read as a heuristic rather than a proven universal law.
Durable rents sit in bottlenecks and trust layers. New arXiv evidence shifts value toward recursive training data flywheels, state management, synthetic environments, and verifier infrastructure.
The next frontier looks less like manually prompt-engineered multi-agent systems and more like models that natively learn when to recurse, what to remember, and what to verify.
The Shift
From thin agent wrappers to recursive training data flywheels, state management, synthetic environments, and verifier infrastructure. The next frontier looks less like manually prompt-engineered multi-agent systems and more like models that natively learn when to recurse, what to remember, and what to verify.