pith. sign in

arxiv: 2604.23072 · v1 · submitted 2026-04-24 · 💻 cs.AI

Analytica: Soft Propositional Reasoning for Robust and Scalable LLM-Driven Analysis

Junyan Cheng , Kyle Richardson , Peter Chin This is my paper

Pith reviewed 2026-05-08 11:30 UTC · model grok-4.3

classification 💻 cs.AI
keywords LLM agentsSoft Propositional Reasoningbias reductionvariance reductionforecasting tasksagent architecturerobust linear modelsJupyter Notebook agent
0
0 comments X

The pith

Analytica reframes LLM analysis as estimating soft truth values of propositions to cut bias and variance through decomposition and linear synthesis.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models struggle with unstable reasoning on tasks like financial forecasting because of systematic bias in how they interpret facts and random variation in their outputs. Analytica tackles this by treating analysis as the estimation of soft truth values for outcome propositions, which can be formally broken down into bias and variance components. The system decomposes problems into trees of subpropositions, assigns tool-using LLM agents to ground and score each one, then recombines the results with robust linear models that average away stochastic noise. This yields measurable gains in accuracy and stability on economic, financial, and political forecasting benchmarks. If the approach holds, it offers a more verifiable and scalable way to apply LLMs to high-stakes real-world analysis.

Core claim

Analytica introduces Soft Propositional Reasoning as a structured process of estimating soft truth values for outcome propositions, allowing formal modeling of estimation error in terms of bias and variance. It operationalizes this through a parallel divide-and-conquer framework that decomposes problems into subproposition trees, employs tool-equipped LLM grounder agents including a Jupyter Notebook agent for data validation, and recursively synthesizes grounded leaves with robust linear models that average out stochastic variance while supporting interactive what-if analysis.

What carries the argument

Soft Propositional Reasoning (SPR), which models analysis as estimation of soft truth values and minimizes error by parallel decomposition to reduce bias plus linear synthesis to reduce variance.

If this is right

  • On economic, financial, and political forecasting tasks Analytica improves accuracy 15.84 percent on average over diverse base models.
  • With a Deep Research grounder it reaches 71.06 percent accuracy and the lowest observed variance of 6.02 percent.
  • The Jupyter Notebook grounder variant delivers 70.11 percent accuracy while cutting cost by 90.35 percent and time by 52.85 percent.
  • Performance remains stable and grows near-linearly with analysis depth, showing resilience to added noise.
  • The architecture adapts to open-weight LLMs and extends to scientific domains beyond forecasting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bias-variance framing could be tested on non-forecasting tasks such as hypothesis generation in scientific literature review.
  • Interactive what-if synthesis might be combined with existing agent toolkits to support dynamic scenario planning in policy or investment settings.
  • If grounding quality improves with better tools, the linear synthesis step may allow scaling to deeper decomposition trees without proportional variance growth.
  • The approach suggests a general template for controlling stochastic error in other multi-step LLM pipelines that currently rely on prompting alone.

Load-bearing premise

That systematic decomposition into subpropositions combined with tool-based grounding by LLM agents can reliably reduce bias, and that robust linear models can average out stochastic variance without introducing new fitting artifacts or losing signal.

What would settle it

A new set of forecasting tasks where Analytica fails to improve accuracy by at least 5 percentage points or shows higher variance than the base LLM models when using the same grounders would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.23072 by Junyan Cheng, Kyle Richardson, Peter Chin.

Figure 1
Figure 1. Figure 1: Given a complex query (e.g., forecasting $NVDA), Analytica selects the most plausible view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of estimation vari￾ance and bias. Analytica with a linear rule has lower bias (closer to the ground truth of 1) and variance. Hitting a better trade-off. LLM Agents for Real-world Analysis We also fo￾cus on the growing body of work using LLM agents to tackle a wide range of open-ended analysis tasks, such as societal dynamics (Cheng & Chin, 2024a), fi￾nancial forecasting (Yu et al., 2024), … view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of Analytica. First, in the view at source ↗
Figure 6
Figure 6. Figure 6: This forms the basis for the strategy of employing an Analyzer to achieve a detailed break view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy vs. number of nodes view at source ↗
Figure 6
Figure 6. Figure 6: Performance vs. cost trade-off analysis. The plots visualize accuracy against monetary view at source ↗
Figure 7
Figure 7. Figure 7: Gradient surfaces of a simple logic formula view at source ↗
Figure 8
Figure 8. Figure 8: Noise sensitivities of a simple logic formula view at source ↗
Figure 9
Figure 9. Figure 9: A Gantt chart illustrating the timespans of the predictive market events included in our view at source ↗
Figure 10
Figure 10. Figure 10: Statistical significance of the results in Table 2, computed by a pairwise McNemar’s view at source ↗
Figure 11
Figure 11. Figure 11: Consensus matrix of final predictions across all methods. The color of each cell rep view at source ↗
Figure 12
Figure 12. Figure 12: Cost-efficiency analysis of different model combinations for the Analytica components. view at source ↗
Figure 13
Figure 13. Figure 13: Marginal impact of model choice on component performance. The chart shows the view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of task correctness rates across all models for different categories. The view at source ↗
Figure 15
Figure 15. Figure 15: Correlation between task features and model performance. The boxplots show that higher view at source ↗
Figure 16
Figure 16. Figure 16: An example of the resynthesis feature for “what-if” scenario analysis. An analyst man view at source ↗
Figure 17
Figure 17. Figure 17: Distribution of the learned weights (βj ) and intercept (β0) for the Linear synthesis rule. The concentration of weights at low positive values demonstrates the rule’s noise-dampening prop￾erty, as formally proven in § A.1. Linear Rule The stability of the Linear rule, P = β0+ PβjCj , is predicated on its ability to act as a weighted average that dampens noise from its inputs. Our experiments confirm that… view at source ↗
read the original abstract

Large language model (LLM) agents are increasingly tasked with complex real-world analysis (e.g., in financial forecasting, scientific discovery), yet their reasoning suffers from stochastic instability and lacks a verifiable, compositional structure. To address this, we introduce Analytica, a novel agent architecture built on the principle of Soft Propositional Reasoning (SPR). SPR reframes complex analysis as a structured process of estimating the soft truth values of different outcome propositions, allowing us to formally model and minimize the estimation error in terms of its bias and variance. Analytica operationalizes this through a parallel, divide-and-conquer framework that systematically reduces both sources of error. To reduce bias, problems are first decomposed into a tree of subpropositions, and tool-equipped LLM grounder agents are employed, including a novel Jupyter Notebook agent for data-driven analysis, that help to validate and score facts. To reduce variance, Analytica recursively synthesizes these grounded leaves using robust linear models that average out stochastic noise with superior efficiency, scalability, and enable interactive "what-if" scenario analysis. Our theoretical and empirical results on economic, financial, and political forecasting tasks show that Analytica improves 15.84% accuracy on average over diverse base models, achieving 71.06% accuracy with the lowest variance of 6.02% when working with a Deep Research grounder. Our Jupyter Notebook grounder shows strong cost-effectiveness that achieves a close 70.11% accuracy with 90.35% less cost and 52.85% less time. Analytica also exhibits highly noise-resilient and stable performance growth as the analysis depth increases, with a near-linear time complexity, as well as good adaptivity to open-weight LLMs and scientific domains.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Analytica, an LLM agent architecture based on Soft Propositional Reasoning (SPR). SPR decomposes complex analysis into trees of subpropositions, grounds them via tool-equipped LLM agents (including a novel Jupyter Notebook agent for data-driven tasks), and synthesizes results using robust linear models to reduce bias and variance. Theoretical modeling of estimation error is claimed, along with empirical results on economic, financial, and political forecasting tasks showing 15.84% average accuracy improvement over diverse base models, peak accuracy of 71.06% with 6.02% variance using a Deep Research grounder, and 70.11% accuracy with 90.35% cost and 52.85% time reductions using the Jupyter agent. Additional claims include noise resilience, near-linear time complexity, and adaptability to open-weight LLMs.

Significance. If the accuracy and stability gains can be isolated to the SPR decomposition and linear synthesis steps, Analytica would offer a structured, bias-variance-aware approach to improving LLM reliability in forecasting and analysis tasks. The introduction of the cost-effective Jupyter Notebook grounder and the reported performance growth with analysis depth are concrete strengths that could influence agent design in applied domains.

major comments (2)
  1. [Abstract] Abstract: The reported 15.84% accuracy improvement is measured against 'diverse base models,' but the abstract does not state whether these baselines are provided with equivalent tool-equipped LLM grounder agents (Deep Research or Jupyter Notebook). Without this control, the gains cannot be attributed to SPR subproposition decomposition or robust linear synthesis rather than richer external data access and code execution, which directly undermines the central claim that SPR plus linear averaging drives the improvement.
  2. [Abstract] Abstract: The paper states that SPR reframes analysis to 'formally model and minimize the estimation error in terms of its bias and variance,' yet supplies no equations, derivations, or explicit bias-variance decomposition. This absence makes it impossible to verify that the robust linear models reduce variance without introducing fitting artifacts or losing signal, as asserted in the synthesis step.
minor comments (2)
  1. [Abstract] Abstract: Performance figures (71.06% accuracy, 6.02% variance, 70.11% with Jupyter) are given without mention of the number of trials, statistical tests, or dataset details, which would strengthen the noise-resilience and stability claims.
  2. [Abstract] Abstract: The term 'Soft Propositional Reasoning (SPR)' and the 'Jupyter Notebook grounder agent' are introduced without formal definitions or pointers to related prior work on propositional structures in reasoning systems.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the presentation of our contributions. We provide point-by-point responses to the major comments below and indicate the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The reported 15.84% accuracy improvement is measured against 'diverse base models,' but the abstract does not state whether these baselines are provided with equivalent tool-equipped LLM grounder agents (Deep Research or Jupyter Notebook). Without this control, the gains cannot be attributed to SPR subproposition decomposition or robust linear synthesis rather than richer external data access and code execution, which directly undermines the central claim that SPR plus linear averaging drives the improvement.

    Authors: We agree that the abstract does not explicitly clarify the baseline configuration. The diverse base models refer to standard LLM agents without the tool grounders or SPR structure. To strengthen the paper, we will revise the abstract to state this clearly and add experiments comparing against tool-equipped baselines without the SPR and linear synthesis components. This will help attribute the gains more precisely to our proposed architecture. revision: yes

  2. Referee: [Abstract] Abstract: The paper states that SPR reframes analysis to 'formally model and minimize the estimation error in terms of its bias and variance,' yet supplies no equations, derivations, or explicit bias-variance decomposition. This absence makes it impossible to verify that the robust linear models reduce variance without introducing fitting artifacts or losing signal, as asserted in the synthesis step.

    Authors: The referee correctly notes the absence of explicit equations in the manuscript. We will add a formal bias-variance decomposition along with key derivations to the revised version, either in the abstract or in a dedicated theoretical subsection. This will substantiate the claims regarding error minimization through the soft propositional framework and linear synthesis. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical measurement

full rationale

The paper introduces Analytica via Soft Propositional Reasoning (SPR) as a divide-and-conquer architecture that decomposes problems into subpropositions, grounds them with tool-equipped LLM agents, and synthesizes via robust linear models to reduce bias and variance. The headline accuracy gains (15.84% average lift, 71.06% peak) are presented strictly as measured empirical outcomes on forecasting tasks rather than as quantities derived by construction from any fitted parameter, self-referential definition, or self-citation chain. No equations appear in the provided text that equate the reported performance to the inputs, and the central claims remain falsifiable against external baselines. The derivation is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Only the abstract is available, so the ledger captures the high-level premises stated there; full paper may introduce additional fitted parameters or assumptions.

axioms (1)
  • domain assumption LLM reasoning errors can be usefully decomposed into bias and variance components that are independently reducible by decomposition and averaging.
    Stated as the principle underlying SPR and the error-minimization strategy.
invented entities (2)
  • Soft Propositional Reasoning (SPR) no independent evidence
    purpose: Reframe complex analysis as estimation of soft truth values of propositions to enable formal bias-variance modeling.
    New framing introduced to structure the agent architecture.
  • Jupyter Notebook grounder agent no independent evidence
    purpose: Provide data-driven fact validation and scoring within the decomposition tree.
    Novel tool-equipped component claimed to improve grounding.

pith-pipeline@v0.9.0 · 5615 in / 1476 out tokens · 74740 ms · 2026-05-08T11:30:34.237407+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 1 internal anchor

  1. [1]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    URLhttps://api.semanticscholar.org/CorpusID:236493564. 12 Published as a conference paper at ICLR 2026 Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. Danny ...

    work page internal anchor Pith review arXiv 2026

  2. [2]

    Matter-of-Fact: A Benchmark for Verifying the Feasibility of Literature-Supported Claims in Materials Science

    doi: 10.18653/v1/2025.emnlp-main.203. URLhttps://aclanthology.org/2025. emnlp-main.203/. Peter Jansen, Oyvind Tafjord, Marissa Radensky, Pao Siangliulue, Tom Hope, Bhavana Dalvi Mishra, Bodhisattwa Prasad Majumder, Daniel S Weld, and Peter Clark. CodeSci- entist: End-to-end semi-automated scientific discovery with code-based experimentation. In Wanxiang C...

    work page doi:10.18653/v1/2025.emnlp-main.203 2025

  3. [3]

    URLhttps://aclanthology.org/2025

    doi: 10.18653/v1/2025.findings-acl.692. URLhttps://aclanthology.org/2025. findings-acl.692/. Ezra Karger, Houtan Bastani, Chen Yueh-Han, Zachary Jacobs, Danny Halawi, Fred Zhang, and Philip E Tetlock. Forecastbench: A dynamic benchmark of ai forecasting capabilities.arXiv preprint arXiv:2409.19839, 2024. Seth Karten, Wenzhe Li, Zihan Ding, Samuel Kleiner,...

    work page doi:10.18653/v1/2025.findings-acl.692 2025

  4. [4]

    doi: 10.18653/v1/2024.emnlp-main.63

    Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main.63. URL https://aclanthology.org/2024.emnlp-main.63/. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scien- tist: Towards fully automated open-ended scientific discovery.arXiv preprint arXiv:2408.06292, 2024. 13 Published as a conference p...

    work page doi:10.18653/v1/2024.emnlp-main.63 2024

  5. [5]

    Futurex: An advanced live benchmark for llm agents in future prediction.arXiv preprint arXiv:2508.11987, 2025

    URLhttps://arxiv.org/abs/2508.11987. Honghua Zhang, Po-Nien Kung, Masahiro Yoshida, Guy Van den Broeck, and Nanyun Peng. Adapt- able logical control for large language models.Advances in Neural Information Processing Sys- tems, 37:115563–115587, 2024. Ben Zhou, Kyle Richardson, Xiaodong Yu, and Dan Roth. Learning to decompose: Hypothetical question decomp...

    work page arXiv 2024

  6. [6]

    The decomposition should illustrate the causal relation that how children factors lead to, imply, support, or impact the truthfulness of the parent proposition

    A proposition is a single sentence statement, with financial, economic, business, social, and political meaning that can be associated with a boolean value True or 40 Published as a conference paper at ICLR 2026 False. The decomposition should illustrate the causal relation that how children factors lead to, imply, support, or impact the truthfulness of t...

    work page 2026

  7. [7]

    Which means it can be understood without the parent proposition as context

    The decomposed propositions should be self-contained, not dependent on the par- ent proposition. Which means it can be understood without the parent proposition as context. For example, it should not refer to the parent proposition using terms like ”it”, ”this metric”, ”this event”, etc

    work page

  8. [8]

    Instead, it is ideal to decompose the proposition into high-level and meaningful financial, economic, business, social, and political factors, assump- tions, hypotheses, etc

    You are not expected to decompose the proposition into low-level fine-grained propositions. Instead, it is ideal to decompose the proposition into high-level and meaningful financial, economic, business, social, and political factors, assump- tions, hypotheses, etc

    work page

  9. [9]

    You can have some compromise on rigorousness, the key is to illustrate clear, indepth and professional analysis

    You should keep the tree to be in-depth but not redundant, this means that you do not need to create commonsense as a child proposition. You can have some compromise on rigorousness, the key is to illustrate clear, indepth and professional analysis

    work page

  10. [10]

    Think comprehensively, deeply, and professionally

    Try to provide really insightful information from your analysis and the outcome de- composition tree that creates ”alpha” for the user. Think comprehensively, deeply, and professionally. You are encouraged to give a really deep analysis and very deep decomposition tree

    work page

  11. [11]

    p true": <float> # the probability of the proposition being true, between 0 and 1

    Do not make redundant propositions, such as the rewrite of the same proposition or the ones that can be simply derived from the negation of other children. Ideal Decomposition(Example for Linear rule): Ideally, a parent proposition can be represented as a multiple linear combination of its children’s propositions, i.e. P true = beta 0 + beta 1*P true1 + b...

    work page 2026

  12. [13]

    You need to think beyond the given data and provide a more comprehensive, in- depth, and broad analysis especially for the points that might be omitted by the grounders

    work page

  13. [14]

    beta":{{

    It is also your task to check the consistency of the children’s proofs and their Ptrue, as well as the quality of the proofs themselves. System Prompt for Synthesizer (Linear) You are an expert for a team of advanced research agents (grounders) in financial, eco- nomic, business, social, and political analysis. The grounders have access to external databa...

    work page 2026

  14. [16]

    You need to think beyond the given data and provide a more comprehensive, in- depth, and broad analysis especially for the points that might be omitted by the grounders, they are the core factors you need to consider in deriving the intercept, the intercept can be seen as an assumption of those omitted factors and the risk factors, remember to clearly sta...

    work page

  15. [17]

    formula": <string>, # e.g., (P1 AND P2) OR (P3 AND NOT PA)

    The weights do not necessary to be from 0 to 1, it can be any real number, and the intercept beta 0 can be negative, but its absolute value should be less than {abs intercept max}, the final P true after the weights and intercept are applied must be between 0 and 1. Please compute yourself first to make sure the final P true is valid before providing your...

    work page 2026

  16. [18]

    You are encouraged to use the knowledge and theory from academia or industry and cite them in your proof

    work page

  17. [19]

    You need to think beyond the given data and provide a more comprehensive, in- depth, and broad analysis especially for the points that might be omitted by the grounders, they are the core factors you need to consider in deriving the formula, remember to clearly state those assumptions in your proof and explain how they affect the formula

    work page

  18. [20]

    The assumption variable id in the formula should ALW AYS BE ”PA”, and all the other variables in the formula should be the proposition id of the children in the input proposition information

    work page

  19. [21]

    You should use ALL the children propositions in the formula, and the formula should be a valid logical combination of the children’s Ptrue

    work page

  20. [22]

    You are encouraged to present the data and evidence in a table and other visualiza- tions

    work page

  21. [23]

    p true": Your estimated probability (a float between 0.00 and 1.00) that the proposition is true, based on your proof and notebook analysis. 2

    The available operators include AND, OR, NOT and parentheses. E.3 GROUNDER General Grounder Prompt You are an expert in financial, economic, business, social, and political analysis. You will be provided with a proposition, and your task is to provide a comprehensive proof that either proves or disproves the proposition. It should include the bullet point...

    work page 2026

  22. [24]

    Synthesis of Child Propositions • P1 (p true = 0.7895): Positive total shareholder return over 12 months • P2 (p true = 0.9040): Long’s risk-adjusted return ¿ short’s • P3 (p true = 0.9320): Shorting NVDA is materially more expensive • P4 (p true = 0.7550): One-year returns are positively skewed Causality:If NVDA is likely to rise (P1), delivering superio...

    work page

  23. [25]

    Omitted Factors and Intercept (β0 = +0.05) We set a modest+5%baseline bias to capture: • Equity risk premium and upward drift in broad markets (Ibbotson & Chen, 2003) • Management execution, capital allocation optionality (M&A, partnerships) • Secular AI tailwinds beyond near-term catalysts • Regulatory, geopolitical, or macro shocks that affect longs and...

    work page 2003

  24. [26]

    best option

    Weight Allocation We allocate weights based on each factor’s direct bearing on the “best option” judgment and our confidence in the children’s proofs. The weights sum to1−β 0 = 0.95: 48 Published as a conference paper at ICLR 2026 Child Propositionp true Weight (βi) Justification P1 0.7895 0.20 Underpins absolute upside, but influenced by valuation and ma...

    work page 2026

  25. [27]

    Computation ofP true Ptrue ≃β 0 + X βi ·p truei = 0.05 + (0.20·0.7895) + (0.30·0.9040) + (0.30·0.9320) + (0.15·0.7550) = 0.05 + 0.1579 + 0.2712 + 0.2796 + 0.1133 ≃0.8720 (87.2%)

    work page

  26. [28]

    NVIDIA stock is likely to deliver a positive total shareholder return (TSR) over the next twelve months

    Risk Assessment Key risks that could materially lower this probability include: • A sudden, deep global recession or credit-market shock reversing equity drift • Unanticipated, severe export controls on NVIDIA’s China business • Technological leapfrogging by AMD, Intel, or hyperscaler custom silicon • Rapid Fed tightening that compresses high-growth multi...

    work page 2026

  27. [29]

    Data Center revenue of $22.6 B. . . up 427% YoY

    NVIDIA Q1 FY2025 Actuals vs. Q1 FY2024 Baseline Metric Q1 FY2024 Q1 FY2025 YoY Growth Data Center Revenue $4.29 B $22.60 B +427% Compute (GPU) Revenue $3.77 B* $19.00 B*>5×YoY (≈>400%) *Estimates derived from disclosed growth multiples in transcript. Key Evidence • Management: “Data Center revenue of $22.6 B. . . up 427% YoY . . . ” • “Compute revenue gre...

    work page

  28. [30]

    Revenue of $26 B. . . well above our outlook of $24 B

    NVIDIA Q2 FY2025 Guidance vs. Q2 FY2024 Baseline Metric Q2 FY2024 (Actual) Q2 FY2025 (Guidance) Implied YoY Growth Total Revenue $16 B $24 B +50% Key Evidence • Transcript: “Revenue of $26 B. . . well above our outlook of $24 B.” • Baseline Q2 FY2024 revenue was $16 B. • $24 B guidance implies exactly +50% YoY; given the Data Center segment is the fastest...

    work page

  29. [31]

    ND series v5

    Hyperscaler Infrastructure Announcements We surveyed public hyperscaler news (AWS, Azure, GCP) for signs of compute-capacity scaling: •A WSlaunched multiple H100 GPU clusters (P5 instances) in every major region in H1 2024. •Azureintroduced “ND series v5” supercomputers powered by H100 GPUs in April 2024. •GCPexpanded “A3” TPU and “A2” A100 GPU pods with>...

    work page 2024

  30. [32]

    AI infrastructure spending to grow>50%annually

    Third-Party Market Forecasts 53 Published as a conference paper at ICLR 2026 Source Forecast Horizon CAGR / YoY Growth Notes IDC (2023) 2023–2025 56% CAGR “AI infrastructure spending to grow>50%annually.” Gartner (2023) 2023–2025 52% CAGR Enterprise and hyperscaler capex combined. McKinsey (2024) 2023–2025 50%+ YoY Focus on generative AI compute budgets. ...

    work page 2026