Stepmark Partners

The State of AI: 2025 Year in Review

—

by

Da
in

Table of Contents

Introduction

The pace of value creation in AI compressed timelines that once spanned years into quarters. Products launched in early 2025, from deep research capabilities to memory features to agentic coding tools, are already table stakes. Valuations recalibrated sharply, with leading frontier labs reaching $300–500 billion and hyperscaler AI capex exceeding $350 billion in aggregate. Government procurement moved from pilots to workforce-scale deployment. China’s open-weight ecosystem emerged as a credible alternative to Western closed models. This review synthesizes the key developments across macro, technology, frontier labs, open-source, government adoption, and regulation, providing a structured lens on a year that reshaped the competitive landscape before most participants could fully internalize the prior quarter.

I. 2025 in Review: Macroeconomic Backdrop

1.1 Markets and Monetary Policy

  • Monetary policy set the tone for the year. Inflation proved persistent, opening at 3.0% in January and easing only marginally to 2.9% by year-end, remaining stubbornly above the Federal Reserve’s 2% target. 
  • The labor market, meanwhile, showed clear signs of softening. Unemployment rose from 4% in January to 4.4% by September, according to data released in November following delays from the government shutdown. 
  • Payroll growth followed a similar trajectory, slowing from 143,000 jobs in January to 119,000 in September. 

Source: (Harvard Gazette)

  • Sticky inflation and a weakening labor market raised stagflation concerns, placing the Fed’s dual mandate in direct tension. Policymakers faced a difficult balancing act and held rates steady through the first half of the year, despite persistent pressure from the Trump administration to ease and the President’s repeated criticism of Chair Powell as “Mr. Too Late.” The Fed held rates steady through the first half of the year despite persistent pressure from the Trump administration to ease. When policymakers did act, they moved decisively, delivering three consecutive rate cuts totaling 75 basis points across September, October, and December. The most recent reduction on December 10 came with a divided 9-3 vote, underscoring internal debate over the path forward.
  • Trade and fiscal policy introduced additional volatility. On April 2, dubbed “Liberation Day,” the administration unveiled a 10% baseline tariff on most imports, alongside elevated reciprocal rates targeting countries with significant bilateral trade surpluses. Markets sold off sharply: the S&P 500 dropped 4.8% and the Nasdaq fell 6% the following session, wiping out $3.1 trillion in market value in the steepest single-day decline since 2020. 

Source (Center for American Progress)

  • The Department of Government Efficiency (DOGE), led by Elon Musk, aimed to reduce federal spending but saw its efforts largely offset by the “One Big Beautiful Bill Act,” which added over $4 trillion to projected deficits. Musk departed in May, and DOGE was disbanded by November.

1.2 Asset Class Performance

  • Asset class performance reflected the uncertain environment. Gold led the way, surpassing $4,000 per ounce for the first time on October 10 and closing the year up 63%. 
  • Digital assets rallied early on regulatory optimism. On January 17, the $TRUMP meme coin launched on the Solana platform and surpassed a $10 billion market cap within an hour. In March, the administration established a Strategic Bitcoin Reserve, and by October 6, Bitcoin reached a record $124,000. However, risk-off sentiment took hold on October 10, triggering a $19 billion single-day liquidation that sparked a sharp reversal. The selloff wiped $1.2 trillion from the market. The selloff wiped $1.2 trillion from the market. Later in October, on October 23, President Trump pardoned Binance founder Changpeng “CZ” Zhao, reversing his 2024 conviction related to anti-money-laundering compliance and reinforcing a more permissive regulatory environment for digital assets. By year-end, Bitcoin was down 9% YTD.
  • As of December 12th, equities rebounded from the April Liberation Day selloff, with the Nasdaq gaining 20% and the S&P 500 rising 16%, while the 10-year Treasury yield fell 38 basis points.
  • The U.S. dollar weakened sharply, posting its steepest annual decline in three decades. The dollar index fell to 98.6 following the December rate cut, its lowest level since late October, capping a year in which it dropped over 10% through the first half alone—the largest such decline since 1973. The greenback fell against major currencies, including the euro and sterling, as expectations for continued rate cuts into 2026 prompted broad selling pressure.

↑ Back to Table of Contents

II. 2025 in Review: Technology Sector

II. 2025 in Review: Technology Sector

On February 3, 2025, President Trump signed Executive Order 14196 directing the Treasury and Commerce Departments to establish a U.S. sovereign investment vehicle within twelve months. The strategy relied on transaction-based investments rather than surplus funding. On August 22, the administration converted $8.9 billion of CHIPS Act grants into an approximately 9.9% equity stake in Intel. The government also obtained special governance rights in the Nippon Steel–U.S. Steel transaction. These actions aligned the federal capital directly with semiconductor manufacturing and industrial assets designated as strategically important.

(Source: Truth Social)

In January 2025, President Trump announced the $500 billion Stargate Project, backed by OpenAI, SoftBank, Oracle, and Abu Dhabi’s MGX, with $100 billion allocated for initial deployment across five U.S. data-center campuses totaling approximately 10 gigawatts of capacity. 

(Source: BBC News)

(Source: Vulcan Elements) 

2.2 Selected Technology Transactions

In 2025, the U.S. tech and AI landscape was defined by a rapid escalation in strategic capital formation—spanning sovereign-adjacent infrastructure announcements, semiconductor re-shoring commitments, and major platform transactions—alongside a parallel set of policy and commercialization catalysts that shaped labor mobility, autonomy deployment, and the monetization path for the AI supply chain. The common thread was convergence: chips, data centers, networking, and regulatory posture increasingly moved in lockstep with frontier AI deployment.

Key developments:

2.3 Energy and Power for AI

2.4 Hyperscaler Capital Deployment

2.5 Neocloud Infrastructure Buildout

AI infrastructure became a primary arena for scale advantages in 2025, with neoclouds using IPO proceeds, long-dated capacity leases, and vertical integration attempts to lock in power, land, and GPUs. The year’s highest-signal moves centered on CoreWeave’s public-market financing and contract stack, the durability of hyperscale demand signals, and the emergence of alternative capacity providers (including Nebius) as capital formation accelerated across the GPU supply chain.

CoreWeave and adjacent infrastructure:

↑ Back to Table of Contents

III. 2025 in Review: Frontier AI Labs

3.1 OpenAI

M&A and Vertical Integration

In 2025, OpenAI accelerated vertical integration across hardware, product infrastructure, and enterprise software, using acquisitions and minority investments to internalize critical capabilities and extend its platform beyond models into devices and applications.

Key transactions:

Product & Feature Development

OpenAI expanded ChatGPT into a multi-modal, agentic, and monetized platform through a steady cadence of launches, positioning it as both a consumer assistant and an enterprise workflow layer.

Key launches and updates (chronological):

  • o3-mini (January 2025): Small reasoning model, released shortly after DeepSeek R1
  • Tasks (beta) (January 2025): Scheduled and recurring actions for Plus, Team, and Pro users
  • Operator (February 2025): Browser-native autonomous agent for ChatGPT Pro ($200/month) with user approval for sensitive actions
  • Deep Research (March 2025): Pro-tier workflows supporting text, PDFs, images, and spreadsheets
  • Memory (April 2025): Persistent personalization for Plus and Pro users, with user controls
  • GPT-4o update & rollback (May 2025): Rolled back after feedback on tone; follow-up updates added shopping tools (recommendations, images, reviews, purchase links)
  • o3-pro (October 2025): Higher-performance reasoning model; followed by an ~80% price reduction
  • Sora 2 (November 2025): TikTok-style, physics-aware video app with cameos and parental controls; surpassed 1 million downloads in fewer than five days since launch, ranking among the fastest-growing consumer apps of 2025
  • Personality 
  • GPT-5.2 (December 2025): Major upgrades in speed, reasoning depth, and hallucination reduction

Commercial deployments:

Compute & Strategic Partnerships

OpenAI prioritized diversified, long-duration compute access to support model training, inference, and product expansion.

Funding & Capital Structure

OpenAI combined primary capital raises, secondaries, and strategic supplier commitments to fund rapid scale while improving long-term unit economics.

Strategic Positioning & Governance

OpenAI’s governance and economic reset unfolded in distinct steps during the second half of 2025.

Key developments:

↑ Back to Table of Contents

3.2 Anthropic

Funding, Valuation & Capital Structure

Anthropic’s 2025 was defined by rapid capital formation, sharp valuation step-ups, and expanding financial flexibility as it scaled Claude for enterprise and developer use cases.

Key capital and governance events:

Feature Development and Model Evolution

Across 2025, Anthropic’s product roadmap emphasized depth over breadth, with feature development centered on coding performance, agentic reliability, and institutional deployment readiness. Rather than prioritizing consumer-facing functionality, the company layered incremental model upgrades with discrete platform features—such as dedicated coding agents, education tooling, voice interaction, and open agent standards—culminating in best-in-class production models optimized for enterprise and regulated environments.

Key Features and Model Updates (2025)

  • Claude Sonnet 4.1 (January 2025): Incremental gains in coding accuracy, instruction adherence, and latency, reinforcing Sonnet as a stable mid-tier production model.
  • Claude Opus 4.1 (January 2025): Improved complex reasoning and long-context reliability, targeting enterprise and research-grade workloads.
  • Claude Haiku 4.1 (February 2025): Faster inference and lower cost, optimizing the lightweight model for high-volume, low-latency use cases.
  • Claude Voice Mode (March 2025): Introduced multimodal voice interaction with enterprise-grade safety and deployment controls.
  • Claude for Education (April 2025): Launched Learning Mode with institutional privacy and administration, deploying at Northeastern University, the London School of Economics, and Champlain College, with integrations into Canvas and Internet2.
  • Claude Code (May 2025): Released a dedicated coding agent for generation, refactoring, and testing, reaching a $500M annualized run rate with rapid adoption.
  • Claude Sonnet 4.2 (June 2025): Interim upgrade improving code generation, tool use, and structured outputs ahead of Sonnet 4.5.
  • Claude Opus 4.2 (July 2025): Strengthened agentic behavior, tool orchestration, and long-context stability for complex enterprise workflows.
  • Claude Sonnet 4.5 (September 2025): Major upgrade widely viewed as a top-tier production coding model, excelling in generation, refactoring, and testing.
  • Claude Opus 4.5 (November 2025): Flagship release delivering leading performance in complex coding, long-context reasoning, and agentic workflows.
  • Model Context Protocol & Agentic AI Foundation (December 2025): Donated the Model Context Protocol and launched the Agentic AI Foundation to advance open, interoperable, and responsible agentic AI standards.

Compute, Infrastructure & Strategic Partnerships

Anthropic expanded compute access and enterprise distribution through hyperscaler alignment and services partnerships.

Key partnerships and infrastructure moves:

Legal, Regulatory & Platform Risk

Anthropic navigated several high-profile legal and operational moments as Claude’s adoption expanded.

Key developments:

↑ Back to Table of Contents

3.3 Google Gemini

Product & Model Development

Google’s 2025 product cadence centered on re-architecting Search around AI-native experiences while scaling Gemini across consumer, developer, and enterprise workflows.

Key launches and feature updates:

  • Gemini in Workspace (January 2025): Included Gemini by default in Workspace Business and Enterprise plans, removing standalone AI add-on pricing.
  • AI Overviews & AI Mode (March 2025): Upgraded AI Overviews with Gemini 2.0 and introduced AI Mode, an AI-first Search experience for complex, multi-step queries.
  • AI Mode U.S. Rollout (May 2025): Expanded AI Mode across the U.S., marking a shift from link-based Search toward AI-driven discovery.
  • Gemini App Integrations (May 2025): Deployed Gemini natively across Gmail, Calendar, Docs, Sheets, and Slides as a built-in productivity layer.
  • Gemini 3 (November 2025): Launched Google’s most advanced model, scoring 37.4 on Humanity’s Last Exam and reaching 650M monthly users and 13M developers.
  • Nano Banana Pro / Gemini 3 Pro Image (November 2025): Released a high-end image generation and editing model for consumer and enterprise workflows.
  • NotebookLM Video Overviews (November 2025): Added Video Overviews (“AI Slides”), enabling narrated, slide-based summaries generated from user-provided sources.

Deals, Talent & Developer Ecosystem

Google’s developer-tooling strategy accelerated following the collapse of OpenAI’s attempted acquisition of Windsurf, creating an opening for alternative deal structures.

Search Distribution, Chrome & U.S. Antitrust

Google’s AI expansion unfolded alongside significant U.S. legal rulings reshaping Search distribution and platform leverage.

Key regulatory developments:

Why the Browser Became Strategic

Control of the browser has historically been one of the most defensible positions in technology, requiring global distribution, default placement, deep OS integration, and massive ongoing security and performance investment. As AI agents, search, and commerce converge, the browser has re-emerged as a critical control point—triggering regulatory scrutiny, acquisition interest, and new AI-native entrants despite the high barriers to entry.

Education as an Early Distribution Battleground

By 2025, education emerged as a strategic distribution wedge for AI platforms rather than a near-term revenue vertical. Model developers increasingly targeted students as a high-leverage cohort—users with high daily engagement, low acquisition costs, strong peer-to-peer diffusion, and a natural conversion path into paid professional tiers. The goal was early habit formation: embedding AI into studying, research, writing, and coding workflows during formative years.

Across vendors, education strategies converged around subsidized or default access, tools optimized for learning artifacts rather than raw answers, and deep integration into institutional systems such as productivity suites and learning management platforms. Monetization was intentionally deferred, with education positioned as a long-term funnel for durable distribution, brand trust, and lifetime value.

Company Actions (2025)

OpenAI

  • Student promotion (Mar–May 2025): Offered two free months of ChatGPT Plus to verified U.S. and Canadian college students, seeding premium feature usage ahead of conversion.
  • Positioning: Framed ChatGPT as a study companion across writing, coding, exam prep, and research workflows.

Google (Gemini & NotebookLM)

  • Gemini for Students: Launched a student-focused Gemini experience emphasizing studying, summarization, brainstorming, and preparation. Offers free 1-year Gemini 3 Pro access with advanced AI tools and 2 TB storage.
  • NotebookLM for Education: Marketed NotebookLM as an AI study tool, converting user-provided notes and PDFs into summaries, study guides, mind maps, and narrated slide explanations (“AI Slides”).
  • Distribution leverage: Scaled adoption through Google Workspace for Education, bypassing incremental purchasing decisions.

Anthropic

  • Claude for Education (April 2025): Introduced Learning Mode to guide reasoning step-by-step rather than deliver direct answers.
  • Institutional embedding: Early campus deployments (Northeastern, LSE, Champlain) with Canvas and Internet2 integrations.
  • Strategic posture: Prioritized pedagogy alignment and institutional trust over viral growth.

Perplexity

  • Student access programs: Provided free or extended Perplexity Pro access to students through verified and campus-based initiatives.
  • Campus Partner Program: Granted up to one year of free Pro to student ambassadors, driving organic adoption via referral loops.
  • Core use case: Positioned as a citation-aware AI research engine for academic work.

Strategic Implication

Education became the earliest point at which AI platforms could establish default workflows and long-term user affinity. Competitive advantage increasingly hinged not on model quality alone, but on who could embed AI most seamlessly into student routines—shaping usage patterns that would persist into professional and enterprise contexts.

↑ Back to Table of Contents

3.4 Meta

Product & Model Development

Meta’s 2025 AI roadmap combined large-scale consumer deployment, generative media innovation, and a strategic pivot from open-source models toward proprietary systems.

  • AI Business Agents Expansion (April 2025): Expanded AI-powered business agents across WhatsApp, Instagram, and Facebook, embedding AI into messaging, commerce, and creator workflows.AI Business Agents Expansion (April 2025): Expanded AI-powered business agents across WhatsApp, Instagram, and Facebook, embedding AI into messaging, commerce, and creator workflows.
  • Llama 4 Voice Capabilities and Behemoth Reassessment (June 2025): Prioritized voice-first, real-time interaction in Llama 4 Scout and Maverick, while internal performance concerns delayed Behemoth and prompted reassessment of the open-model strategy.
  • Vibes AI Video Feed (September 2025): Launched Vibes, an AI-native video feed within the Meta AI app for creating and sharing short-form, AI-generated videos.
  • AI Recommendations and Surface Enhancements (October 2025): Improved AI-driven recommendations and expanded AI surfaces across Meta apps, deepening integration into core experiences.
  • Avocado Closed Model Disclosure (December 2025): Disclosed development of Avocado, a closed, monetizable next-generation AI model signaling a strategic shift from open-source Llama, with reporting indicating training pipeline elements from Alibaba’s Qwen system; on the day the news surfaced, Meta shares fell approximately 1.2% while Alibaba shares rose roughly 2%.
  • AI Chatbot–Driven Ad Targeting (December 2025): Announced use of Meta AI chatbot interactions for ad targeting, excluding sensitive categories and eliminating opt-outs for engaged users.

Talent, Organization & AGI Push

Meta paired product expansion with aggressive M&A and infrastructure investment to support its AI ambitions.

Talent, Organization & AGI Push

Meta’s 2025 talent strategy included high-impact executive hires and broad technical recruitment to accelerate its AI roadmap.

  • Meta Superintelligence Labs Formation (June 2025): Mark Zuckerberg announced the creation of Meta Superintelligence Labs, consolidating advanced AI research, product, and infrastructure efforts under a unified superintelligence strategy.
  • Safe Superintelligence Engagement & Daniel Gross Hire (June 2025): Following discussions that did not result in an acquisition of Safe Superintelligence, Meta recruited co-founder Daniel Gross,  co-founder of Safe Superintelligence into its advanced AI leadership ranks to lead frontier intelligence initiatives, reflecting continued efforts to attract top research talent amid intensifying competition across leading AI labs.
  • Senior AI Research Hiring Surge (June to July 2025): Meta executed a broad AI recruiting push, hiring at least eight senior researchers from leading labs, including at least 4 from OpenAI, with publicly identified hires such as Shengjia Zhao, Jiahui Yu, Shuchao Bi, and Hongyu Ren.
  • Ruoming Pang Hired from Apple (July 2025): Ruoming Pang, former head of Apple’s AI foundation models team, left Apple to join Meta in a high-profile move reportedly backed by compensation exceeding $200 million.
  • Andrew Tulloch Departure and Re-Hire (October 2025): Co-founder of Thinking Machines Lab, who rejected a compensation offer reportedly reaching $1B in August 2025, Andrew Tulloch was reported by The Wall Street Journal in October 2025 to be departing the startup to join Meta, as revealed in an internal message announcing his exit.
  • Yann LeCun Departure (November 2025): Yann LeCun, Meta’s longtime Chief AI Scientist and founding leader of FAIR, announced plans to depart after 12 years to co-found a new AI research company focused on next-generation intelligence beyond current LLM paradigms.
  • Soumith Chintala Departure (November 2025): Soumith Chintala, PyTorch cofounder and long-time leader of Meta’s core AI infrastructure, departed Meta to join Mira Murati’s Thinking Machines Lab amid the startup’s senior hiring push.
  • Apple Design Leadership Migration (December 2025): Meta hired Alan Dye, Apple’s longtime head of Human Interface Design, as Chief Design Officer to lead AI and Reality Labs product and interaction design. The move followed broader Apple leadership transition planning under Tim Cook and included additional senior hires such as Billy Sorrentino (former Director of iOS/macOS Design) and Ke Yang (Apple’s AI web search lead), signaling Meta’s renewed focus on consumer-facing AI experiences.
  • Meta Talent Strategy (2025): Meta’s aggressive hiring push included $100M+ signing bonuses for top AI researchers, as part of a broader effort to attract leading AI talent amidst intensifying competition with OpenAI and Google. 
  • AI-Native Hiring Practices (2025): Meta announced it would allow candidates to use AI assistants during coding interviews, aligning hiring evaluation with real-world developer workflows as AI agents increasingly write, refactor, and test production code (“vibecoding”).

Legal, Regulatory & Market Signals

Meta navigated rising regulatory scrutiny while securing strategic legal wins in 2025.

↑ Back to Table of Contents

3.5 xAI

In 2025, xAI’s strategy centered on vertical integration, distribution expansion, and sovereign-scale infrastructure, combining X to internalize data and reach, extending Grok distribution via major platforms, and securing power and compute through large-scale partnerships.

Key Developments

Product & Feature Development

xAI pushed a rapid cadence of flagship releases paired with premium packaging, while navigating recurring safety and trust challenges as Grok scaled across mass-consumer platforms.

Key launches and updates:

  • Grok 3 (February 2025): Launched as xAI’s next flagship model, marketed as trained on ~200,000 GPUs, with deep research, image generation, and “Big Brain” reasoning modes; distributed via X Premium+ and a standalone SuperGrok tier.
  • Premium Monetization Structure (February 2025): Grok distribution emphasized paid access via X Premium+, a $30/month SuperGrok tier, and an ultra-premium ~$300/month plan, reinforcing xAI’s positioning around power-user monetization rather than broad freemium access.
  • Grok 4 (July 2025): Released shortly before/alongside heightened controversy following harmful outputs, reinforcing that model behavior can become a platform-level risk at consumer scale.
  • Grok 4 Fast (September 2025): Introduced as a lower-latency / lower-cost variant, positioned around materially improved inference efficiency for search-style tasks and scaled deployment.
  • Tesla integration (December 2025): Expanded via Tesla’s Holiday Update, enabling Grok-driven voice navigation workflows (e.g., adding/editing multi-stop routes), tightening the xAI–Tesla product loop beyond a standalone chatbot.

Commercial packaging signals:

  • Premium tiering (2025): Grok monetization emphasized paid access (X Premium+ and standalone tiers), including an ultra-premium $300/month plan launched alongside Grok 

Funding & Capital Structure

xAI blended primary capital formation with secondaries and debt to fund GPU procurement, site expansion, and retention economics in an escalating talent war.

Key financings and liquidity events:

Legal, Regulatory & Competitive Posture

As xAI scaled aggressively, its competitive posture increasingly spilled into the courtroom, reflecting rising tensions around AI distribution, talent mobility, and market access.

Key legal developments:

  • Antitrust Litigation Against Apple and OpenAI (August 2025): xAI sued Apple and OpenAI, alleging exclusionary practices tied to ChatGPT distribution and platform integration, raising broader antitrust questions around default placement and AI market access.
  • Trade Secret & Talent Poaching Lawsuit Against OpenAI (September 2025): xAI separately accused OpenAI of trade secret misappropriation and coordinated employee poaching tied to Grok development, escalating an already adversarial competitive relationship.

↑ Back to Table of Contents

IV. 2025 in Review: Open-Source and Open-Weight Models

Before the emergence of ultra-large “Behemoth” models, Meta’s LLaMA series represented the leading Western open-weight foundation model. However, the April 2025 release of LLaMA 4 did not deliver a clear step-change in performance, weakening its role as the default open base layer. At the same time, U.S. chip restrictions pushed Chinese labs toward efficiency-first scaling, emphasizing MoE architectures, systems optimization, and inference-cost engineering. Against this backdrop, China’s open-source ecosystem—spanning big tech, startups, and research institutions—advanced rapidly, competing on performance per unit of compute and accelerating adoption across models, tools, and downstream applications.

4.1 Key 2025 moments

↑ Back to Table of Contents

V. 2025 in Review: Government Adoption

In 2025, government adoption of frontier AI models shifted from isolated pilots to standardized procurement and workforce-scale deployment. Defense, civilian agencies, and allied governments increasingly treated large language models as strategic infrastructure—using prototype awards, catalog-based purchasing, and promotional pricing to accelerate evaluation, reduce procurement friction, and establish multi-vendor competition across mission-critical use cases.

  • Anthropic–United Kingdom Government Memorandum of Understanding (February 2025): Anthropic signed a memorandum of understanding with the United Kingdom government to explore Claude deployment in public services, working with the United Kingdom Artificial Intelligence Security Institute on safety, secure infrastructure, and labor-market impact analysis.
  • OpenAI for Government Launch and Department of Defense Prototype Award (June 2025): OpenAI launched OpenAI for Government alongside a Department of Defense (DoD) Chief Digital and Artificial Intelligence Office (CDAO) prototype award with a $200M ceiling covering frontier AI use cases across defense and enterprise domains.
  • Department of Defense Multi-Vendor Agentic AI Awards (July 2025): The Department of Defense (DoD) Chief Digital and Artificial Intelligence Office (CDAO) issued additional $200M-ceiling prototype awards to Anthropic, Google, and xAI, formalizing a four-vendor cohort for agentic AI evaluation.
  • General Services Administration OneGov — ChatGPT Enterprise (August 2025): The General Services Administration (GSA) added ChatGPT Enterprise to the OneGov catalog at $1 per federal agency for a one-year promotional evaluation term.
  • General Services Administration OneGov — Anthropic Claude (August 2025): The General Services Administration (GSA) added Anthropic Claude at $1 per agency across all three branches of government.
  • General Services Administration OneGov — Gemini for Government (August 2025): The General Services Administration (GSA) added Gemini for Government at $0.47 per agency under a promotional term cited through September 30, 2026.
  • General Services Administration OneGov — xAI Grok (September 2025): The General Services Administration (GSA) added xAI Grok at $0.42 per agency, with the offer cited as valid through March 2027.
  • Department of Defense GenAI.mil Workforce Rollout (December 2025): The Department of Defense (DoD) selected Gemini for broad unclassified deployment via GenAI.mil, signaling the shift from procurement experimentation to workforce-level rollout.
  • Pricing Clarification — “Per Agency” Model (2025): OneGov pricing reflects symbolic, government-wide access fees per participating agency for evaluation periods rather than per-seat pricing or normalized commercial annual recurring revenue.

↑ Back to Table of Contents

VI. 2025 in Review: Regulatory Developments

6.1 United States

The U.S. regulatory environment in 2025 combined executive action, state-level statutes, export controls, and selective enforcement tied directly to AI model deployment and training.

  • Expanded AI chip export controls to China (January 2025): The U.S. Department of Commerce implemented new export control rules restricting shipments of advanced AI accelerators to China while easing licensing pathways for allied countries, materially reshaping global GPU supply chains.
  • Federal adoption of frontier AI models (June 2025): The Department of Defense awarded OpenAI a prototype contract with a ceiling of two hundred million dollars, embedding frontier AI models into defense and enterprise government workflows.
  • Copyright and training data enforcement escalation (July 2025): Content owners moved aggressively to monetize AI training data through licensing and litigation, highlighted by The New York Times licensing arrangement with Amazon and Reddit’s legal action against Anthropic over alleged scraping violations.
  • California SB 53 frontier AI transparency law (September 2025): California passed legislation requiring large frontier model developers to disclose safety protocols and report serious AI related incidents, introducing state level oversight of advanced model risk management.
  • California SB 243 on AI companion chatbots (October 2025): California enacted SB 243, establishing safety obligations for AI companion systems used by minors, including disclosures, safeguards against harmful content, and enforcement penalties.
  • Federal executive order on AI and state preemption (December 2025): President Trump issued an executive order directing federal agencies to review and challenge state AI laws deemed burdensome to innovation or interstate commerce, reinforcing a federal first approach to AI governance.

6.2 United Kingdom and European Union

Across the UK and EU, 2025 policy focused on coupling safety oversight with compute access, research funding, and industrial scale up.

6.2.1 United Kingdom

  • UK Anthropic memorandum of understanding (February 2025): The UK government signed a formal agreement with Anthropic to collaborate on frontier model safety research and evaluation, reinforcing the UK’s role as a global AI safety hub.
  • Frontier AI regulatory framework development (March 2025): The UK government advanced its frontier AI safety framework centered on model evaluations, testing, and voluntary commitments, positioning regulatory access as a strategic advantage for AI developers.
  • AI research and compute subsidy expansion (May 2025): The UK expanded public funding for AI research infrastructure and evaluation institutes, using subsidized compute access rather than prescriptive regulation to influence model development.
  • Parliamentary push for binding AI legislation (December 2025): Cross party lawmakers publicly called for mandatory regulation of powerful AI systems, signaling growing pressure to convert voluntary frameworks into enforceable law.

6.2.2 European Union

  • EU AI Act implementation phase (January 2025): The European Union entered the operational implementation phase of the AI Act, advancing compliance obligations for general purpose and high risk AI systems including documentation, transparency, and post market monitoring.
  • AI Continent Action Plan announcement (March 2025): The European Commission advanced an industrial strategy linking AI regulation with large scale investment in data centers, advanced chips, and AI gigafactories to retain compute capacity within the bloc.
  • Cloud and AI Act proposal (June 2025): The Commission introduced a complementary proposal focused on scaling AI compute while managing energy, grid, and climate constraints, explicitly tying AI expansion to infrastructure policy.
  • DMA and DSA enforcement affecting AI distribution (Throughout 2025): Enforcement actions against designated gatekeepers increasingly shaped how AI models were bundled, distributed, and surfaced across major digital platforms.

6.3 China

China’s AI governance in 2025 emphasized centralized oversight, content transparency, and strict service compliance across the AI stack.

  • Measures for labeling AI generated content issued (March 2025): Chinese regulators released mandatory rules requiring explicit and implicit labels on AI-generated text, images, audio, and video.
  • AI generated content labeling rules effective (September 2025): The labeling measures entered into force nationwide, requiring platforms and developers to implement watermarking and disclosure mechanisms at scale.
  • Generative AI service compliance framework reinforced (Throughout 2025): China continued enforcement of registration, security review, and transparency requirements for generative AI services, integrating model oversight into national cybersecurity and data governance regimes.
  • Deep synthesis and algorithm regulation enforcement (Throughout 2025): Existing rules governing recommendation algorithms and deep synthesis technologies remained actively enforced, mandating transparency, opt-out mechanisms, and safeguards against harmful or misleading content.

↑ Back to Table of Contents

A Year to Remember

2025 will be remembered as the year AI transitioned from technological promise to operational reality; slowly at first, then suddenly all at once. What began as incremental progress accelerated into a cascade of breakthrough moments: open-source models achieved performance parity with their closed counterparts, fundamentally democratizing access to frontier capabilities, while agentic AI systems moved beyond demonstrations to execute complex, multi-step workflows in production environments. 

The sheer velocity of product announcements, infrastructure buildouts, and commercial deployments this year underscored an industry no longer constrained by the question of “if,” but animated by the urgency of “how fast” and “at what scale.” 

For enterprises navigating M&A opportunities, and for anyone tracking the trajectory of technology, the inflection point is unmistakable: AI is reshaping competitive dynamics, business models, and strategic priorities across every sector. 

As we close out this remarkable year, we look ahead with conviction that the momentum will only intensify. In January, we will share our predictions for 2026, a year that will determine which business models endure, which applications reach mainstream adoption, and where the next wave of value creation will emerge.