Not investment advice. This page is for informational and educational purposes only. Sterling is not a registered investment advisor. Nothing here is a recommendation to buy, sell, or hold any security. Markets carry risk and you may lose money. Do your own research and consult a licensed professional before making any investment, securities purchase, or trading decision.

Thesis

Why this sector matters to investors right now. Structural, not market timing.

AI software in mid-2026 is a two-track story. The first track is large and visible: hyperscaler capex on the underlying compute reached approximately 358 billion dollars in 2025 across Microsoft, Alphabet, Meta, Amazon, and Oracle, with consensus tracking ~445 billion in 2026 (cash-flow capex; aggregated from company 10-Ks and quarterly disclosures plus JPM, MS, GS sell-side consensus). Frontier-lab annualized revenue has scaled in step. OpenAI reached approximately 13 billion dollars in annualized run-rate by late 2025 (The Information, Reuters). Anthropic reached approximately 5 billion (The Information, CNBC). Microsoft's Azure AI services contributed roughly 13 percentage points to Azure-and-other-cloud growth in FY2025-Q4 (Satya Nadella, earnings call). Google Cloud processed approximately 480 trillion tokens per month at I/O 2025, up roughly 50 times year over year, then more than doubled again to roughly 1,300 trillion tokens per month exit-rate by Q3 2025 (Sundar Pichai, Google I/O 2025 and Alphabet Q3 2025 earnings).

The second track is the margin and durability question. The marginal cost of frontier-equivalent intelligence has fallen roughly 24 times since the original GPT-4 list price (30 dollars per million input tokens, March 2023) to GPT-5 (1.25 dollars per million input tokens, August 2025), with open-weights models (DeepSeek-V3, Llama 4 Maverick, Qwen3 235B) sitting at or below 0.30 dollars per million input tokens at comparable capability (Artificial Analysis pricing history; provider list prices). DeepSeek-R1 (January 2025) compressed the closed-versus-open Chatbot Arena Elo gap to roughly 50 to 100 Elo points from a 200-plus gap in early 2024 (LMSys Chatbot Arena leaderboard, October 2025 snapshot). The investible question is no longer whether the technology works at the frontier. It is which layer of the stack captures durable margin as capex inflates and inference prices compress.

Structural drivers

Forces that shape long-run demand and economics. Each driver is sourced.

•Hyperscaler capex is rising sharply and majority-AI-attributable. Combined Microsoft, Alphabet, Meta, Amazon, and Oracle cash-flow capex grew from approximately 240 billion in 2024 to approximately 358 billion in 2025, with 2026 consensus at approximately 445 billion. Each of Mark Zuckerberg, Andy Jassy, and Larry Ellison has stated on the record that the majority (or vast majority) of incremental capex is AI infrastructure. Source: company 10-Ks and quarterly cash-flow disclosures; JPMorgan, Morgan Stanley, Goldman Sachs sell-side consensus for 2026E.
•Frontier-lab revenue is compounding rapidly. OpenAI annualized run-rate moved from approximately 1.6 billion (Q4 2023) to approximately 13 billion (late 2025), more than tripling year over year. Anthropic moved from approximately 0.1 billion (Q4 2023) to approximately 5 billion (mid-to-late 2025). xAI moved from approximately 0.1 billion (Q4 2024) to approximately 0.8 billion (late 2025). Source: The Information, Reuters, Bloomberg reporting on internal projections and management disclosures.
•Token volume is exploding at the platform layer. Google disclosed 480 trillion tokens per month at I/O 2025 (May 2025), up roughly 50 times year over year, and more than doubled to roughly 1,300 trillion exit-rate by Q3 2025. Microsoft Azure AI Foundry processed approximately 100 trillion tokens in calendar Q1 2025 (about 33 trillion per month), up 5 times year over year. OpenAI cited approximately 6 billion tokens per minute peak API throughput in Q3 2025. Source: Google I/O 2025 keynote, Microsoft Build 2025, OpenAI DevDay 2025 and Brad Lightcap commentary.
•Enterprise adoption is broad and rising. McKinsey's State of AI 2025 survey reported 78 percent of organizations using AI in at least one business function, up from 72 percent the prior year and 33 percent in 2023 (gen-AI specifically). Per-function gen-AI use roughly doubled across marketing and sales, service operations, product development, and engineering and IT from 2023 to 2024. Source: McKinsey 'The state of AI in 2025' (Nov 2025) and prior annual surveys.
•ChatGPT and Copilot adoption are concrete enterprise proof points. ChatGPT weekly active users reached approximately 800 million in late 2025, up from 300 million (Aug 2024) and 500 million (Mar 2025). Microsoft said nearly 70 percent of the Fortune 500 are using Microsoft 365 Copilot and paid seats nearly tripled year over year as of FY2025-Q1. Salesforce reported over 5,000 paid Agentforce deals by Q1 FY2026. ServiceNow Now Assist Pro Plus net new ACV grew triple-digit percent year over year for five consecutive quarters into Q3 2025. Source: OpenAI corporate disclosures; MSFT, CRM, NOW earnings calls.
•Agentic and multimodal use cases are expanding the addressable surface. Salesforce's Agentforce launched late 2024 and reached the thousands-of-paid-deals milestone within two quarters. Anthropic's Claude Sonnet 4.5 (Sep 2025) and Opus 4 (May 2025) drove coding-agent volumes at Cursor (~200 million ARR mid-2025 per The Information) and GitHub Copilot. Multimodal flagships (GPT-4o, Gemini 2.5 Pro, Llama 4 Maverick with 10M context) expanded the supported input modalities. Source: company earnings, The Information reporting, provider blog posts.
•Capital availability remains very high for top labs. 2025 AI VC dollar volume reached approximately 205 billion (PitchBook, CB Insights), more than double 2024's approximately 100 billion. Mega-rounds dominate the mix: OpenAI 40 billion at 300 billion valuation (Mar 2025, SoftBank-led), Anthropic raises totaling 13 billion-plus across 2025, xAI ~10 billion at 50 billion (May 2025), Databricks 10 billion-plus follow-on. Source: Reuters, Bloomberg, The Information, Anthropic and Mistral company press releases.
•Stargate-class infrastructure commitments anchor multi-year compute supply. OpenAI, SoftBank, Oracle, and MGX announced Stargate in January 2025: a stated 500 billion dollars over four years in US AI infrastructure, with 100 billion targeted for immediate deployment, first site Abilene, Texas. Source: OpenAI Stargate announcement, January 2025.

Structural risks

Forces that could compress demand, change economics, or break the thesis.

•Open-weights commoditization is real and accelerating. DeepSeek-V3 (Dec 2024) and DeepSeek-R1 (Jan 2025) reached GPT-4-class capability at sub-0.30 dollars and sub-1 dollar per million input tokens respectively, with weights freely downloadable under permissive licenses. Llama 4 Maverick (Apr 2025) and Qwen3 235B (Apr 2025) closed the Chatbot Arena Elo gap to within roughly 50 to 100 points of closed flagships at the October 2025 snapshot, versus a 200-plus gap in early 2024. The closed-API margin thesis is structurally weaker than it was 18 months ago. Source: Artificial Analysis pricing history; LMSys Chatbot Arena leaderboard October 2025 snapshot.
•API price decay is brutal at the frontier capability tier. GPT-4 launched at 30 dollars per million input tokens (March 2023). GPT-5 launched at 1.25 dollars per million input tokens (August 2025), a roughly 24 times decline in 29 months. GPT-4o mini and DeepSeek-V3 sit at 0.15 to 0.27 dollars per million input tokens at GPT-4-class capability, roughly a 100 to 200 times cut. Token revenue scales with volume but per-token margin is compressing fast. Source: Artificial Analysis pricing history; OpenAI, Anthropic, DeepSeek pricing pages.
•Regulatory phase-in continues. EU AI Act prohibitions took effect February 2, 2025; GPAI obligations applied from August 2, 2025; high-risk system obligations apply from August 2, 2026; embedded-product obligations apply from August 2, 2027 (Regulation 2024/1689). The systemic-risk GPAI threshold sits at 10 to the 25th FLOPs (the EU AI Act anchor), which captures most frontier models. Colorado AI Act took effect February 1, 2026 and is the first comprehensive US state AI consumer-protection law. California AB 2013 (training-data disclosure) took effect January 2026. Source: EU Official Journal 2024/1689; European Commission AI Office; Colorado General Assembly SB 24-205; California AB 2013.
•Copyright and licensing litigation is unresolved at scale. NYT v. OpenAI and Microsoft (S.D.N.Y. 1:23-cv-11195), Authors Guild v. OpenAI, Getty v. Stability AI, and RIAA v. Suno and Udio are all in discovery or pre-trial stages as of mid-2026. A plaintiff-friendly outcome on any of training-data fair use, output regurgitation, or statutory damages would materially raise the cost basis for frontier training. Source: Court Listener case dockets; RIAA complaints.
•Capex digestion risk if AI revenue conversion lags. Combined hyperscaler 2026 capex of approximately 445 billion implies revenue scaling that has to materialize. Microsoft, Alphabet, Amazon, and Meta together generate ~13 percent of Azure growth (MSFT only) plus undisclosed AI contribution at the other three. The free-cash-flow profile of the four has tightened materially as capex outpaces operating cash flow on AI-specific assets. A multi-quarter air pocket in AI revenue versus capex would compress every layer of the stack simultaneously. Source: company 10-Ks; CNBC Feb 6, 2026 and Fortune April 30, 2026 reporting cited in adjacent AI Chips sector entry.
•Talent costs are inflating. Meta's June 2025 acquisition of 49 percent of Scale AI at an implied roughly 29 billion dollar valuation (~14.3 billion paid) was widely characterized as a talent-and-data acquisition that brought Alexandr Wang to Meta as Chief AI Officer. Microsoft's March 2024 acqui-hire of Inflection AI paid approximately 650 million in license fees for the team. Top-tier ML researcher compensation packages crossed eight figures during 2024 to 2025 (The Information reporting). Source: The Information; press releases from Meta, Microsoft.
•Concentration risk on a small number of compute and customer relationships. OpenAI is heavily dependent on Microsoft Azure plus the announced Stargate buildout (OpenAI plus SoftBank plus Oracle plus MGX). Anthropic compute is concentrated at AWS (8 billion-plus combined commitment) and Google (2 billion-plus). xAI is concentrated at Oracle Cloud Infrastructure plus owned facilities. Disruption to any one of these relationships would have outsized effects on the corresponding lab. Source: company press releases; Reuters, Bloomberg coverage.
•China policy and export-control volatility. BIS export controls on H100, H200, and B200 to China remain in force; the AI Diffusion Framework (January 2025) was paused under the Trump administration. The Cyberspace Administration of China requires algorithm registration and security assessment for any public-facing generative AI service. US-China decoupling on the model stack remains an open variable that affects DeepSeek, Qwen, Hunyuan, and Ernie distribution outside their home market. Source: BIS Federal Register notices; CAC Interim Measures (Aug 2023).
•Public-market AI sentiment is two-sided. DeepSeek-R1 (Jan 20, 2025) triggered a sharp NVDA selloff on cost-collapse fears. CoreWeave's March 2025 IPO priced at 40 dollars per share, below the marketed range, and Cerebras IPO has been pending since its September 2024 S-1 filing. The market is willing to fund mega-rounds at frontier labs but discriminates sharply against narrower or higher-cyclicality AI infrastructure issuers. Source: CoreWeave IPO release March 27, 2025; Reuters DeepSeek market reaction reporting.

Competitive landscape

How to think about the players. Framing along axes (pure play vs diversified, incumbent vs challenger, etc). Not stock picking.

The investible universe sorts into six archetypes, each with very different economics. NVIDIA, AMD, Broadcom, TSMC, and other compute-layer names are cross-referenced to the AI Chips sector and not analyzed here.

1. Frontier model labs (closed). OpenAI (approximately 13 billion run-rate late 2025, 300 billion March 2025 SoftBank-led primary valuation), Anthropic (approximately 5 billion run-rate, 61.5 billion March 2025 primary valuation), xAI (approximately 0.8 billion run-rate, 50 billion reported May 2025 valuation), Google DeepMind (inside Alphabet, no standalone disclosure), Cohere (5.5 billion July 2024 Series D, regulated-enterprise focus), AI21, Reka. Economics: heavy capex spend, sticky enterprise contracts at the top, exposed to API price decay and open-weights substitution at the bottom of their model SKU range.

2. Frontier model labs (open-weights). Meta Llama family (Llama 4 Maverick/Scout/Behemoth, April 2025), Mistral (6 billion June 2024 Series B), DeepSeek (no external funding, High-Flyer parent), Alibaba Qwen, Tencent Hunyuan, Baidu Ernie, Black Forest Labs, Stability AI. Economics: limited direct monetization, but each release compresses API prices system-wide. Open-weights commoditization is a structural feature, not a temporary disruption.

3. Big Tech AI divisions. Microsoft (Azure AI plus Copilot, approximately 13 billion AI run-rate exit FY2025-Q2 per Nadella, growing 175 percent year over year), Alphabet (Vertex AI plus Gemini API plus Workspace AI plus 480T-then-1,300T tokens per month, no standalone AI revenue line), Amazon (Bedrock plus Q plus AWS generative AI services, multi-billion-dollar run-rate per Jassy without precise dollar figure), Meta (AI embedded in ad ranking and Meta AI consumer app, no separate AI revenue), Apple (Apple Intelligence, no disclosed revenue), Oracle (OCI Generative AI, AI RPO bookings driving FY2025 backlog), IBM (watsonx, 5 billion-plus cumulative book-of-business across FY2024-FY2025 consulting plus software). Economics: AI is a contributor to existing distribution; not yet standalone revenue at most.

4. AI-first cloud and inference platforms (neoclouds). CoreWeave (CRWV, March 2025 IPO at 23 billion valuation), Nebius Group (NBIS, post-Yandex carve-out direct listing October 2024 at 2.6 billion), Applied Digital (APLD), IREN, plus the private layer (Together AI, Fireworks, Lambda Labs, Crusoe). Economics: GPU-fleet leverage, contracted backlog with frontier labs and hyperscalers, capex-heavy. Less differentiated than the hyperscalers but with cleaner AI-attributable revenue lines.

5. Data, ML, and enterprise AI platforms. Databricks (62 billion December 2024 Series J, approximately 3 billion-plus ARR mid-2025), Snowflake (Cortex AI), MongoDB (Atlas Vector Search plus Voyage AI acquisition February 2025), Hugging Face, Scale AI (49 percent acquired by Meta June 2025 at implied ~29 billion), Palantir (AIP, approximately 800 million AI revenue 2025E, US Commercial growing approximately 70 percent year over year), ServiceNow (Now Assist), Salesforce (Agentforce, 5,000-plus paid deals by Q1 FY2026), Adobe (Firefly plus AI Assistant, approximately 700 million AI revenue 2025E). Economics: existing distribution plus AI attach; the most direct enterprise-AI proof points.

6. Application companies. ChatGPT and Claude.ai (lab first-party), Cursor (Anysphere, approximately 200 million ARR mid-2025, 9 billion June 2025 Series C), Perplexity (~100 million ARR mid-2025, 18 billion May 2025 Series F), Glean (7.2 billion June 2025 Series F), Harvey (legal, ~50 million-plus ARR mid-2025), Sierra (4.5 billion October 2024 Series B), Decagon, ElevenLabs, Midjourney (~500 million ARR exit 2024). Economics: highly differentiated by workflow, lower compute concentration, exposed to substitution if the underlying model layer commoditizes faster than enterprise switching cost rises.

Cross-cutting framing: closed labs carry the steepest revenue ramp but the steepest margin-compression risk. Open-weights labs do not directly monetize at scale but materially shape API pricing for everyone. Big Tech AI divisions earn the most absolute AI revenue but the least transparent disclosure. Neoclouds are GPU-fleet leverage plays. Enterprise platforms with existing distribution (Salesforce, ServiceNow, Microsoft) are the cleanest enterprise-AI revenue proxies. Pure-play applications carry the highest sensitivity to the underlying model stack. The right cross-stack question is which layer holds margin in a world where the cost of intelligence keeps falling and aggregate token volume keeps rising.

Key metrics to watch

The operational and financial metrics that matter most in this sector. Each one names its source and update cadence.

Metric	Source	Frequency	Why it matters
Hyperscaler combined cash-flow capex (Microsoft, Alphabet, Meta, Amazon, Oracle)	Quarterly 10-Q filings and earnings calls; consolidated by JPMorgan, Morgan Stanley, Goldman Sachs sell-side research	Quarterly	Combined 2025 capex approximately 358 billion, 2026 consensus approximately 445 billion. A guide-down across multiple hyperscalers in a single cycle would be the cleanest top-down signal that the AI capex super-cycle is normalizing.
Frontier-lab annualized run-rate revenue (OpenAI, Anthropic, xAI)	The Information, Reuters, Bloomberg reporting on internal projections and management disclosure	Episodic; major leaks roughly each quarter	OpenAI 13 billion, Anthropic 5 billion, xAI 0.8 billion as of late 2025. Pace of revenue growth versus pace of model release determines whether closed-lab margin holds.
Microsoft Azure AI services contribution to Azure growth	MSFT quarterly earnings calls (Satya Nadella commentary on points of contribution)	Quarterly	The single cleanest hyperscaler-AI disclosure. Held at approximately 13 percentage points from FY2024-Q4 through FY2025-Q4. A change in this number is a directional read on AI demand at the biggest enterprise-AI vendor.
Google token volume processed across products and APIs	Alphabet earnings calls and Google I/O keynotes (Sundar Pichai disclosures)	Roughly twice yearly; I/O plus selected earnings calls	480 trillion tokens per month at I/O 2025 (up 50x year over year), more than doubled by Q3 2025 to approximately 1,300 trillion. The single largest disclosed token-volume number in the industry and the cleanest proxy for AI consumption growth.
Enterprise AI seat attach: Microsoft 365 Copilot, Salesforce Agentforce, ServiceNow Now Assist	MSFT, CRM, NOW quarterly earnings calls and IR materials	Quarterly	The cleanest enterprise-AI adoption proxies. Microsoft cites Fortune 500 adoption and paid-seat growth; Salesforce cites paid Agentforce deal counts (5,000-plus by Q1 FY2026); ServiceNow cites Pro Plus net new ACV (triple-digit growth for five consecutive quarters into Q3 2025).
Chatbot Arena Elo gap between top closed and top open-weights models	LMSys Chatbot Arena leaderboard (live), with periodic dated snapshots	Daily (raw); analyzed quarterly	October 2025 snapshot: top closed model 1410 Elo (GPT-5), top open-weights 1357 (DeepSeek-R1). Gap roughly 50 to 100 Elo versus 200-plus in early 2024. If the gap closes to within 25 Elo, closed-API margin pressure intensifies.
API price per million input tokens at GPT-4-class capability	Provider list prices (OpenAI, Anthropic, Google, DeepSeek, Together AI hosting Llama and Mistral); Artificial Analysis pricing history	Event-driven, roughly each major model release	Frontier capability dropped from 30 dollars per million input tokens (GPT-4, March 2023) to 1.25 dollars (GPT-5, August 2025), and to sub-0.30 dollars at open-weights tier. The slope of this curve sets the ceiling on per-token revenue.
AI VC funding and frontier-lab valuation marks	PitchBook, CB Insights, Crunchbase; company funding announcements	Quarterly synthesis plus event-driven	2025 AI VC dollar volume approximately 205 billion versus approximately 100 billion in 2024. Concentration is extreme: strip out the top five rounds and the underlying deal pace looks closer to 2024 levels. Watch primary-versus-secondary tender spreads for signals of overheating.

Catalysts and milestones

Known upcoming events that could move the sector. Dated where possible.

•EU AI Act high-risk system obligations apply August 2, 2026. Conformity assessments, risk management systems, data governance, and post-market monitoring become enforceable for AI used in hiring, credit, education, critical infrastructure, law enforcement, and biometrics. Source: EU Regulation 2024/1689; European Commission AI Office.
•EU AI Act embedded-product obligations apply August 2, 2027. Final phase of the multi-year compliance ramp; extends high-risk obligations to AI inside regulated products (medical devices, machinery, vehicles). Source: EU Regulation 2024/1689.
•DeepSeek and other Chinese open-weights frontier releases. DeepSeek-R1 (January 2025) compressed the closed-versus-open Arena Elo gap. Subsequent DeepSeek, Qwen, and Hunyuan releases through 2026 are the cleanest leading indicator of further open-weights pressure on closed-API pricing. Source: model release pages; LMSys leaderboard.
•OpenAI for-profit restructuring outcome. The transition from capped-profit to a fully for-profit structure (subject to negotiation with the OpenAI nonprofit, Microsoft, and state attorneys general in California and Delaware) remains open. Each disclosed milestone changes the cap-table and the regulator posture. Source: OpenAI corporate disclosures; Reuters, The Information reporting through 2025 to 2026.
•Databricks IPO. Repeatedly anticipated, not yet filed. Would be the largest single AI-software public-market test. Source: Databricks press releases; sector S-1 monitoring.
•Stargate site buildout milestones. 500 billion stated four-year US AI infrastructure plan announced January 2025 by OpenAI, SoftBank, Oracle, and MGX, with 100 billion targeted near-term. First site Abilene, Texas. Buildout pace is the most concrete operational signal of frontier compute supply through 2026 to 2028. Source: OpenAI Stargate announcement, January 2025.
•Hyperscaler 2027 capex guidance during 2026 earnings cycle. The 2026 approximately 445 billion combined number is the baseline; 2027 guidance either confirms (extension of the supercycle) or moderates (normalization) the trajectory. Source: MSFT, GOOGL, AMZN, META, ORCL Q3 and Q4 2026 earnings.
•Major copyright litigation rulings. NYT v. OpenAI (S.D.N.Y. 1:23-cv-11195), Authors Guild v. OpenAI, RIAA v. Suno and Udio. Substantive rulings on training-data fair use or output regurgitation would set the precedent for licensing-cost economics across the sector. Source: Court Listener case dockets.
•Frontier-model releases through 2026 and 2027. GPT-5 (Aug 2025) and Claude Sonnet 4.5 plus Opus 4.5 (Sep 2025) are the current frontier; each successor release tests both capability ceiling and pricing. Source: provider release pages.
•Anthropic and OpenAI compute commitments. Anthropic has committed to up to 1 million Google TPUs (CNBC November 2025) and remains a major AWS Trainium2 customer. OpenAI's Stargate commitment and Microsoft Azure exclusivity are similarly multi-year. Each commitment locks in a layer of the stack. Source: CNBC, OpenAI, Anthropic disclosures.

What would change the view

Conditions or evidence that would invalidate the thesis or materially shift the risk picture.

•Open-weights matching closed on Chatbot Arena and major benchmarks simultaneously. The October 2025 gap is 50 to 100 Elo. A close to within 25 Elo (or open-weights overtaking on subsets like math or code) would compress closed-API gross margins materially and accelerate enterprise switching to self-hosted or hosted-open inference.
•Plaintiff-favorable training-data ruling. A clear ruling against fair-use defenses in NYT v. OpenAI or Authors Guild v. OpenAI that requires retroactive licensing for training data would raise frontier-model cost basis. Estimates of incremental licensing costs vary materially; we treat them as a not-yet-knowable range.
•US antitrust action against Microsoft-OpenAI structural relationship. The relationship is unusual: Microsoft has approximately 49 percent of OpenAI's for-profit subsidiary by some reporting, exclusive Azure provision, and substantial commercial overlap. An FTC, DOJ, or EU enforcement action requiring structural separation would reset the AI software competitive map.
•Hyperscaler capex pause or reversal. If Microsoft, Alphabet, Amazon, or Meta materially guides down 2027 capex versus the 2026 approximately 445 billion baseline, the supercycle narrative weakens immediately for every layer of the AI software stack.
•Frontier-lab unit economics turning structurally negative. If publicly cited gross-margin metrics on OpenAI or Anthropic deteriorate materially as inference prices fall faster than volume rises, the revenue-scaling thesis hits a ceiling. Estimates here are noisy and we flag them as such.
•Major US state AI regulation comparable to the EU AI Act high-risk regime. The Colorado AI Act took effect February 2026 but is narrower than the EU regime. A California or New York follow-on of comparable scope would materially raise US compliance costs.
•A China-side compute or model embargo. US export controls and possible EU equivalents are the active state. A reciprocal Chinese restriction on Qwen, Hunyuan, or DeepSeek availability outside China, or a Western embargo on hosting Chinese open-weights models, would re-fragment the open-weights map.
•OpenAI for-profit restructuring blocked or materially restructured. The negotiation outcome with the OpenAI nonprofit, Microsoft, and state attorneys general remains genuinely uncertain. A blocked transition would constrain OpenAI's ability to raise capital at the pace required by its compute commitments.

What we are not covering

Sub-areas, technologies, or companies we are deliberately excluding from the analysis, and why.

•AI semiconductor compute (NVIDIA, AMD, Broadcom, TSMC, Samsung, SK Hynix, Micron, ASML). Covered in the separate AI Chips sector at /sectors/ai-chips. These names sit at a different layer of the stack with different economics (foundry oligopoly, HBM oligopoly, lithography monopoly) and are the principal beneficiaries of the hyperscaler capex flowing through this sector.
•Defense AI (Anduril, Palantir government, Shield AI). Covered in the separate Defense sector. Palantir commercial AIP appears here as a horizontal enterprise-AI platform; the government and defense-tech contracts are out of scope.
•AI in humanoid robotics (Tesla Optimus, Figure, 1X, Apptronik, Agility Robotics). Covered in the separate Humanoids sector. The AI Software analysis here covers foundation models and inference platforms; physical-embodiment robotics is a separate capital and product cycle.
•AI in autonomous driving (Waymo, Tesla FSD, Pony.ai, WeRide, Aurora, Mobileye). Covered in the separate Autonomy sector. Driving-specific AI stacks share frontier-model lineage but have distinct unit economics (rides per vehicle, geofence, regulatory regime).
•AI cybersecurity (CrowdStrike AI features, Palo Alto Cortex, Wiz, SentinelOne, Robust Intelligence). Out of scope for this sector and will live in a future Cybersecurity sector. Flagged here for adjacency only.
•Quantum computing as a substitute or successor to classical AI compute. Different time horizon and technology stack. Adjacent but not part of the current AI software thesis.
•AI used as a marketing label without material AI revenue contribution. We exclude by judgment, not by hard percentage threshold; the criterion is that AI must be material to the investment thesis (a meaningful contributor to revenue growth, product strategy, or competitive position).
•Consumer-only AI hardware (Apple Intelligence, AI PC categories, smart glasses). Apple Intelligence is a descriptive callout in this sector with no disclosed revenue. AI PCs are part of the broader PC cycle and not analyzed here.

Sources

Primary sources cited in this analysis. Links open in a new tab.

Microsoft 10-K and 10-Q filings; FY2025 and FY2026 earnings calls (Nadella, Hood)Accessed 2026-05-15
Microsoft FY2025-Q2 earnings call (Azure AI 13-point contribution; $13B AI run-rate)Accessed 2026-05-15
Microsoft Build 2025 keynote (Azure AI Foundry 100T tokens Q1 2025)Accessed 2026-05-15
Alphabet 10-K and 10-Q filings; quarterly earnings calls (Pichai, Porat)Accessed 2026-05-15
Google I/O 2025 keynote (480T tokens per month disclosure)Accessed 2026-05-15
Amazon 10-K and 10-Q filings; quarterly earnings calls (Jassy on Bedrock and AWS AI)Accessed 2026-05-15
Meta Platforms 10-K and 10-Q filings; Zuckerberg Q4-2024 'majority of capex on AI infrastructure'Accessed 2026-05-15
Oracle 10-K and 10-Q filings; FY2025-Q4 earnings call (Ellison on OCI AI training capex)Accessed 2026-05-15
Apple 10-K and 10-Q filingsAccessed 2026-05-15
IBM 10-K and 10-Q filings (watsonx book-of-business commentary)Accessed 2026-05-15
Salesforce FY2026-Q1 earnings call (Benioff: 5,000-plus paid Agentforce deals)Accessed 2026-05-15
ServiceNow Q3-2025 earnings call (McDermott: five quarters of triple-digit Now Assist growth)Accessed 2026-05-15
Adobe 10-K and 10-Q filings (Firefly and AI Assistant commentary)Accessed 2026-05-15
Palantir 10-K and 10-Q filings (AIP and US Commercial commentary)Accessed 2026-05-15
OpenAI corporate communications and pricing pagesAccessed 2026-05-15
OpenAI GPT-5 launch and pricingAccessed 2026-05-15
OpenAI DevDay 2025 (800M weekly ChatGPT users; 4M developers)Accessed 2026-05-15
OpenAI Stargate project announcement (January 2025)Accessed 2026-05-15
Reuters: OpenAI completes 40 billion funding round at 300 billion (March 31, 2025)Accessed 2026-05-15
Anthropic press release: Series E at 61.5 billion post-money (March 3, 2025)Accessed 2026-05-15
Anthropic Claude Sonnet 4.5 announcement (September 2025)Accessed 2026-05-15
Anthropic Claude 4 family announcement (May 2025)Accessed 2026-05-15
Bloomberg: xAI raising 10 billion at 50 billion valuation (May 29, 2025)Accessed 2026-05-15
DeepSeek-V3 release (December 2024)Accessed 2026-05-15
DeepSeek-R1 release (January 20, 2025)Accessed 2026-05-15
Meta Llama 4 announcement (April 2025)Accessed 2026-05-15
Mistral AI Series B (June 2024)Accessed 2026-05-15
Cohere Series D (July 2024)Accessed 2026-05-15
Databricks Series J at 62 billion (December 2024)Accessed 2026-05-15
Anysphere (Cursor) Series C at 9 billion (June 2025)Accessed 2026-05-15
Glean Series F (June 2025)Accessed 2026-05-15
The Information: Meta acquires 49 percent of Scale AI (June 2025)Accessed 2026-05-15
The Information: Microsoft-Inflection AI license fee (March 2024)Accessed 2026-05-15
CoreWeave IPO pricing release (March 27, 2025)Accessed 2026-05-15
EU AI Act official text (Regulation 2024/1689)Accessed 2026-05-15
European Commission Digital Strategy on AI regulatory frameworkAccessed 2026-05-15
European Commission AI Office (GPAI Code of Practice)Accessed 2026-05-15
US Executive Order 14179 (rescinds EO 14110; January 23, 2025)Accessed 2026-05-15
Colorado AI Act (SB 24-205)Accessed 2026-05-15
California AB 2013 training-data disclosure lawAccessed 2026-05-15
BIS export controls on advanced computing itemsAccessed 2026-05-15
Court Listener: NYT v. OpenAI and Microsoft (case 1:23-cv-11195)Accessed 2026-05-15
Court Listener: Authors Guild v. OpenAI (case 1:23-cv-08292)Accessed 2026-05-15
LMSys Chatbot Arena leaderboardAccessed 2026-05-15
Artificial Analysis (model benchmarks, pricing, latency)Accessed 2026-05-15
McKinsey State of AI annual survey (2024 and 2025)Accessed 2026-05-15
PitchBook AI and machine learning researchAccessed 2026-05-15
CB Insights AI funding reportsAccessed 2026-05-15
CNBC: Anthropic up-to-1-million Google TPU commitment (November 2025)Accessed 2026-05-15
The Information topics page: OpenAI coverageAccessed 2026-05-15
The Information topics page: Anthropic coverageAccessed 2026-05-15

Audit trail

Record of the last review and what changed. Required on every refresh.

Last reviewed: 2026-05-15

Change log

2026-05-15Initial Strategy authoring. All ten required SOP components populated using sourced 2025 full-year and 2026 year-to-date data. Primary sources: company 10-K and 10-Q filings (MSFT, GOOGL, AMZN, META, ORCL, IBM, CRM, NOW, ADBE, PLTR, AAPL); The Information, Reuters, Bloomberg, CNBC reporting on private-lab revenue and rounds; Google I/O 2025, Microsoft Build 2025, OpenAI DevDay 2025 keynotes; Artificial Analysis pricing history; LMSys Chatbot Arena leaderboard October 2025 snapshot; McKinsey State of AI 2024 and 2025 surveys; EU Regulation 2024/1689 and European Commission AI Office; PitchBook and CB Insights AI funding trackers; Court Listener case dockets for NYT v. OpenAI and Authors Guild v. OpenAI; OpenAI Stargate January 2025 announcement; CoreWeave March 2025 IPO release; Anthropic, OpenAI, Mistral, DeepSeek model release pages and pricing pages. All sources accessed 2026-05-15.

Unresolved questions

•OpenAI for-profit restructuring outcome. Negotiation continues with the OpenAI nonprofit board, Microsoft, and California and Delaware state attorneys general. Each disclosed milestone changes the cap table, the Microsoft commercial relationship, and the regulator posture. Watch for material disclosures through 2026 to 2027.
•EU AI Act high-risk rule impact on US labs serving EU customers. August 2, 2026 is the enforcement date. Compliance interpretation by the EU AI Office, the Code of Practice on GPAI, and any early enforcement actions will determine whether the cost is incremental compliance overhead or a material market-access constraint.
•Antitrust resolution on Microsoft-OpenAI structural relationship. FTC, DOJ, and EU posture remains under review. A formal enforcement action requiring structural separation, a behavioral remedy, or a no-action conclusion will each have very different consequences for the Big Tech AI competitive map.
•Whether the closed-versus-open Chatbot Arena Elo gap closes to within 25 Elo through 2026. October 2025 snapshot: 50 to 100 Elo gap. A close to within 25 Elo would compress closed-API margin materially; an opening above 100 Elo would harden closed-lab pricing power.
•Whether 2027 hyperscaler capex materially exceeds, matches, or moderates from the 2026 approximately 445 billion baseline as Q3 and Q4 2026 earnings disclose forward views.
•Databricks IPO timing. Repeatedly anticipated, not yet filed. Would be the largest single AI-software public-market test.
•NYT v. OpenAI and Authors Guild v. OpenAI substantive rulings on training-data fair use and output regurgitation. Plaintiff-favorable rulings would materially raise frontier-model cost basis sector-wide.
•Anthropic and OpenAI compute commitment execution. Anthropic up-to-1-million TPU commitment via Google plus AWS Trainium2 deployment; OpenAI Stargate buildout cadence. Watch for delivered (not just contracted) compute as the binding variable on frontier training pace.