This is not a blog. It is a working archive of what I watch: technology, AI, business, leadership, and the industries being remade in front of us. Each observation records what happened, why it mattered, and what it changed in practice — with sources, so you can check my reading against the record.
Part timeline, part research file, part business analysis. It grows as the world moves.
Featured collection
AI and Business Since 2022
On November 30, 2022, OpenAI released a research preview called ChatGPT. Everything in business — software, support, capital, hiring, regulation — has been adjusting ever since. This collection tracks that adjustment year by year, from the launch through the present.
102 observations · 134 sources · 2022–2026
Start here — six moments that carry the story
· Global · Product launch · Model release
OpenAI launches ChatGPT — the generative AI era begins
OpenAI released ChatGPT as a free research preview — a conversational interface on a GPT-3.5-series model. It reached one million users in five days. For the first time, large language models were usable by anyone, not just researchers and engineers.
It turned years of lab research into a mass-market product overnight and forced every major technology company to respond within weeks. Google declared an internal code red.
Business impact
The fastest product-strategy realignment in modern tech history. Venture funding, enterprise pilots, and the roadmaps of every major software company reorganized around generative AI within a single quarter.
OpenAI released GPT-4 — multimodal, dramatically more capable, and scoring around the top 10% on a simulated bar exam where GPT-3.5 scored near the bottom 10%. Notably, OpenAI disclosed no architecture or training details, citing competition and safety.
GPT-4 reset expectations of what AI could do in professional domains — law, medicine, finance, coding — and became the benchmark every rival measured against for over a year.
Business impact
It powered the first serious knowledge-work deployments, and its closed technical report marked the industry's pivot from open research culture to competitive secrecy.
NVIDIA joins the trillion-dollar club as the H100 boom begins
NVIDIA's Q1 earnings stunned markets with revenue guidance roughly 50% above consensus, driven by data-center GPU demand. Days later its market cap crossed $1 trillion — the first chipmaker ever to reach that level.
It established AI compute as the scarcest strategic resource in technology, and proved the surest profits of the boom were in the hardware layer.
Business impact
H100 allocation became a board-level concern, hyperscaler capex reoriented toward GPU datacenters, and the picks-and-shovels trade defined 2023 markets.
The DeepSeek shock: R1 wipes ~$600B off NVIDIA in a day
DeepSeek open-sourced its R1 reasoning model under MIT license, matching OpenAI o1-level performance at a fraction of the claimed cost. NVIDIA fell 17% — nearly $600 billion, the largest single-day market-cap loss in US history.
Frontier-level reasoning, cheap and open, directly challenged the assumptions underneath hundreds of billions of dollars in US AI capex.
Business impact
Inference prices fell across the industry, open-weight models gained enterprise legitimacy overnight, and US–China AI competition became the market's central theme.
AI reaches the workforce: Amazon cuts 14,000 as the evidence mounts
Amazon announced 14,000 corporate cuts — its largest — citing AI investment. Salesforce's Benioff disclosed AI agents let him cut support from 9,000 to ~5,000 roles. A Stanford study found a 13% relative employment decline for workers aged 22–25 in AI-exposed occupations since late 2022.
MIT: 95% of enterprise GenAI pilots show no P&L impact
MIT NANDA's 'GenAI Divide' report found that despite $30–40 billion in enterprise spending, ~95% of generative AI pilots delivered no measurable P&L impact — attributing failure to an integration learning gap, not model quality. Purchased solutions succeeded about twice as often as internal builds.
The spark. A research preview becomes the fastest-adopted product in history.
· Global · Product launch · Model release
OpenAI launches ChatGPT — the generative AI era begins
OpenAI released ChatGPT as a free research preview — a conversational interface on a GPT-3.5-series model. It reached one million users in five days. For the first time, large language models were usable by anyone, not just researchers and engineers.
It turned years of lab research into a mass-market product overnight and forced every major technology company to respond within weeks. Google declared an internal code red.
Business impact
The fastest product-strategy realignment in modern tech history. Venture funding, enterprise pilots, and the roadmaps of every major software company reorganized around generative AI within a single quarter.
The land grab. Frontier models, hyperscaler alliances, the first enterprise deployments, the first real regulation.
· US · Company move · Funding & markets
Microsoft commits ~$10B to OpenAI
Microsoft announced the third phase of its OpenAI partnership — a multiyear investment widely reported at ten billion dollars. Azure remained OpenAI's exclusive cloud, and Microsoft gained the right to deploy OpenAI models across its entire product line.
It tied the most important AI lab to the second-largest cloud provider and signaled that frontier AI would be financed by Big Tech capital, not standalone startups.
Business impact
Set the template for every hyperscaler–lab alliance that followed (Google and Amazon with Anthropic) and repositioned Microsoft as the perceived AI leader, putting Google's search business under siege.
· Global · Funding & markets · Enterprise adoption
ChatGPT hits 100M users — fastest-growing consumer app in history
A UBS analyst note, citing Similarweb data, estimated ChatGPT reached 100 million monthly active users in January 2023 — two months after launch. TikTok took nine months; Instagram took two and a half years.
Google rushes out Bard; a demo error erases ~$100B of Alphabet
Google announced Bard, its conversational AI built on LaMDA. In the promotional demo, Bard answered a question about the James Webb Space Telescope incorrectly — and Alphabet shares fell nearly 8%, erasing roughly $100 billion in market value in a day.
Microsoft ships AI Bing, then puts Copilot inside Office
Microsoft unveiled an OpenAI-powered Bing and Edge in February, then announced Microsoft 365 Copilot in March — LLMs plus company data inside Word, Excel, PowerPoint, Outlook, and Teams. Nadella's line: AI will fundamentally change every software category, starting with search.
The first time an incumbent shipped frontier AI directly into products used by billions, and an open assault on Google's core business.
Business impact
The 'Copilot' framing became the dominant enterprise AI pattern of 2023, and $30 per user per month set the benchmark for monetizing generative AI in software.
GitHub Copilot goes enterprise as Stack Overflow traffic slides
GitHub made Copilot for Business generally available in February, then announced the GPT-4-powered Copilot X vision in March. Meanwhile Similarweb data showed Stack Overflow traffic falling ~14% month-over-month as developers took their questions to AI instead.
Software development became the first profession with measurable, at-scale AI workflow substitution — visible in the decline of the web's canonical programming Q&A site.
Business impact
AI coding assistants became standard enterprise tooling within a year, and content platforms learned their data was now training-set leverage.
Meta's LLaMA weights leak — open-source AI ignites
Meta released LLaMA to approved researchers under a noncommercial license; within a week the weights leaked via torrent. LLaMA-13B outperformed the much larger GPT-3 on most benchmarks, and the leak put near-frontier weights in anyone's hands.
It ignited the open-source LLM ecosystem — Alpaca, Vicuna, llama.cpp — and the defining open-versus-closed debate of the era.
Business impact
Proved capable models could run on consumer hardware, spawned a parallel open-model economy outside the API giants, and pushed Meta toward formally embracing open release with Llama 2.
The first enterprise wave: Salesforce, Morgan Stanley, Bloomberg
In a single month: Salesforce announced Einstein GPT and a $250M generative AI fund; Morgan Stanley revealed a GPT-4 assistant for its ~16,000 financial advisors; Klarna shipped one of the first ChatGPT plugins; and Bloomberg published BloombergGPT, a 50B-parameter finance-specific model.
Regulated, brand-sensitive industries moved from curiosity to production deployments within four months of ChatGPT's launch.
Business impact
These deals established the enduring enterprise patterns: proprietary data plus foundation model, AI embedded in existing SaaS, and assistant-style distribution.
Salesforce · Morgan Stanley · Bloomberg · Klarna · OpenAI
· Global · Model release
GPT-4: AI passes the bar exam
OpenAI released GPT-4 — multimodal, dramatically more capable, and scoring around the top 10% on a simulated bar exam where GPT-3.5 scored near the bottom 10%. Notably, OpenAI disclosed no architecture or training details, citing competition and safety.
GPT-4 reset expectations of what AI could do in professional domains — law, medicine, finance, coding — and became the benchmark every rival measured against for over a year.
Business impact
It powered the first serious knowledge-work deployments, and its closed technical report marked the industry's pivot from open research culture to competitive secrecy.
Anthropic launches Claude; Claude 2 goes public in July
Anthropic — founded by ex-OpenAI researchers — opened Claude to businesses via API, positioned as helpful, honest, and harmless. In July, Claude 2 launched publicly with a 100K-token context window that could digest hundreds of pages per prompt.
It established a credible safety-focused rival to OpenAI and made long-context document processing a competitive differentiator.
Business impact
Enterprises got a second frontier-model vendor — breaking OpenAI's de facto monopoly on commercial LLM APIs — and Anthropic became the prize in the hyperscaler investment battles later that year.
Midjourney v5 and Adobe Firefly make AI imagery photoreal — and commercial
Midjourney V5 produced the viral 'Pope in a puffer jacket' image that fooled millions — arguably the first mass-scale AI misinformation moment. Days later Adobe launched Firefly, trained on licensed content and built for commercially safe creative work.
AI image generation became simultaneously indistinguishable from photography and packaged for professional use.
Business impact
Agency and stock-photography cost structures began collapsing, and synthetic-media verification became an immediate problem for platforms and newsrooms.
China answers: Baidu's Ernie Bot, Alibaba's Tongyi Qianwen
Baidu unveiled Ernie Bot, China's first major ChatGPT rival — its stock fell 10% during a pre-recorded demo, then rebounded as analysts tested it favorably. Weeks later Alibaba launched Tongyi Qianwen and pledged to integrate it across every Alibaba product.
It opened the parallel Chinese LLM race under a distinct regulatory regime, confirming generative AI as a US–China strategic competition.
Business impact
Chinese cloud and internet giants reoriented their roadmaps around proprietary LLMs, and Beijing followed with the world's first binding generative AI rules that August.
30,000 signatures call for a pause on giant AI experiments
A week after GPT-4, the Future of Life Institute published an open letter calling for a six-month pause on training systems more powerful than GPT-4. Musk, Wozniak, Bengio, and Harari signed. No lab paused.
NVIDIA joins the trillion-dollar club as the H100 boom begins
NVIDIA's Q1 earnings stunned markets with revenue guidance roughly 50% above consensus, driven by data-center GPU demand. Days later its market cap crossed $1 trillion — the first chipmaker ever to reach that level.
It established AI compute as the scarcest strategic resource in technology, and proved the surest profits of the boom were in the hardware layer.
Business impact
H100 allocation became a board-level concern, hyperscaler capex reoriented toward GPU datacenters, and the picks-and-shovels trade defined 2023 markets.
Llama 2: open weights become a sanctioned Big Tech strategy
Meta released Llama 2 free for research and commercial use, with Microsoft as preferred distribution partner on Azure. Unlike the leaked research-only LLaMA, this came with an explicit commercial license.
Open weights went from leak to strategy, giving every company a credible free alternative to paid frontier APIs.
Business impact
Llama 2 became the default foundation for thousands of fine-tuned commercial models, pressured API pricing industry-wide, and set Meta's ecosystem-capture strategy against OpenAI's closed model.
Amazon announced an investment of up to $4 billion in Anthropic. AWS became Anthropic's primary cloud for mission-critical workloads, with Claude offered through Amazon Bedrock and training moving toward Amazon's custom Trainium chips.
Paris-based Mistral AI — months old, backed by a record $113M seed round — released Mistral 7B under Apache 2.0. The 7.3B-parameter model beat Llama 2 13B on every reported benchmark.
A small European startup shipped competitive models months after founding, proving efficient open weights were a viable strategy against US giants.
Business impact
Mistral became Europe's AI champion — central to EU AI Act lobbying and a $2B valuation by December — and its models became workhorses for cost-sensitive deployments.
Executive Order 14110 — the longest in US history — directed over 50 federal entities to act on AI safety, privacy, and security, and invoked the Defense Production Act to require frontier-model developers to report safety-test results to the government.
The US government's first binding intervention in frontier AI development, setting compute-threshold reporting before any legislation existed.
Business impact
Created immediate compliance workstreams at the major labs and federal contractors, and defined the US regulatory posture until its rescission in 2025.
OpenAI · Anthropic · Google · Microsoft · Meta · NVIDIA
· Europe · Regulation & policy
28 nations — including the US and China — sign the Bletchley Declaration
At the first global AI Safety Summit at Bletchley Park, 28 countries plus the EU signed a joint commitment on frontier AI risk — the first time Washington and Beijing acknowledged it together at head-of-state level.
UK Government · OpenAI · Anthropic · Google DeepMind
· US · Product launch · Company move
OpenAI DevDay: GPT-4 Turbo, custom GPTs, and the platform play
OpenAI's first developer conference: GPT-4 Turbo with a 128K context window at roughly a third of GPT-4's input price, custom GPTs with a forthcoming store, and the Assistants API. ChatGPT had reached 100 million weekly active users.
OpenAI pivoted from model vendor to platform company, aiming to own the application layer the way app stores owned mobile.
Business impact
Aggressive price cuts compressed margins across the LLM API market, and the platform strategy reshaped AI startup investment theses overnight — thin wrappers were suddenly endangered.
Hollywood's AI reckoning: SAG-AFTRA wins digital-replica protections
After a 118-day strike in which AI was a central issue, SAG-AFTRA's deal with the studios established the first major AI labor protections: explicit consent and compensation for digital replicas, renewed per project. The WGA had won parallel protections for writers in September.
The first large-scale collective-bargaining settlement over generative AI — consent-and-compensation became the template for AI and human labor.
Business impact
The provisions became reference points across music, gaming, and publishing, signaling that AI cost savings in creative industries would face negotiated limits.
OpenAI's board fires Sam Altman — and reinstates him in five days
OpenAI's nonprofit board ousted its CEO for not being 'consistently candid.' Nearly all ~770 employees threatened to follow him to Microsoft. Five days later Altman returned under a reconstituted board.
The crisis exposed how fragile governance was at the world's most important AI company — and showed that investors, employees, and Microsoft held the real power, not the safety-oriented board.
Business impact
Enterprises accelerated multi-vendor AI strategies to hedge key-person and governance risk, and the episode reframed the debate over how frontier labs should be governed.
Google announced Gemini 1.0 in three sizes — Ultra, Pro, and Nano — claiming Ultra was the first model to beat human experts on MMLU. It was the first major product of the merged Google DeepMind organization.
Google's first credible claimed answer to GPT-4 after a year of being perceived as behind.
Business impact
It restored competitive credibility ahead of Google's Gemini-everywhere strategy, though a staged demo video showed how much scrutiny AI marketing now faced.
EU agrees the AI Act — the world's first comprehensive AI law
After a ~36-hour final negotiation, the European Parliament and Council reached political agreement on the AI Act: a risk-based framework with banned practices, high-risk obligations, and new tiered rules for general-purpose AI.
It set the global regulatory benchmark for AI — the 'Brussels effect' — and covered frontier model providers for the first time anywhere.
Business impact
Companies selling AI into the EU began multi-year compliance programs, with fines up to 7% of global turnover, and the GPAI rules shaped how US labs document and release models worldwide.
The scaling year. Multimodality, reasoning models, the agent turn — and a chipmaker becomes the world's most valuable company.
· US · Model release · Research & science
OpenAI previews Sora: text-to-video arrives
OpenAI unveiled Sora, a diffusion model generating photorealistic video up to a minute long from text prompts. It stayed in research preview, shared only with red-teamers and select creatives.
Google introduced Gemini 1.5 Pro, able to process a million tokens — roughly an hour of video or 700,000 words. A week later Google paused Gemini's people-image generation after it produced historically inaccurate depictions, acknowledging it had overcorrected for diversity.
Long context reset expectations for how much information one model could reason over; the image controversy became the year's cautionary tale on guardrail tuning.
Business impact
Whole-codebase and long-document analysis became practical enterprise use cases, while the image episode made guardrail failure a reputational-risk line item.
Microsoft partners with Mistral, bringing Mistral Large to Azure
Microsoft and Mistral announced a multi-year partnership making Mistral Large available on Azure, with a small Microsoft investment. Europe's leading lab gained global distribution.
Klarna: our AI assistant does the work of 700 agents
Klarna reported its OpenAI-powered assistant handled 2.3 million conversations in its first month — two-thirds of support chats, resolving issues in under two minutes versus eleven, with a projected $40M profit improvement.
Anthropic released the Claude 3 family — Haiku, Sonnet, and Opus — with Opus claiming benchmark leadership over GPT-4 across reasoning, knowledge, and coding, plus vision capabilities.
Cognition launched Devin, billed as the first AI software engineer — an autonomous agent with its own shell, editor, and browser that completed real freelance jobs in demos. GitHub followed in April with Copilot Workspace, a task-oriented AI development environment.
NVIDIA introduced the Blackwell architecture and GB200 systems, claiming up to 30x faster inference and 25x lower cost and energy versus the prior generation, with every major cloud signed up as a first adopter.
Meta launched Llama 3 in 8B and 70B sizes, trained on roughly 15 trillion tokens, claiming best-in-class open-model performance and shipping across every major cloud.
AlphaFold 3 extended structure prediction beyond proteins to their interactions with DNA, RNA, and drug-like molecules, reporting at least 50% better accuracy on molecular interactions, published in Nature with a free research server.
OpenAI launched GPT-4o — one model for text, audio, and vision with voice responses as fast as 232 milliseconds — and made it free to all ChatGPT users at half the API cost of GPT-4 Turbo.
Real-time, natural multimodal interaction reached a mass-market free product, resetting the price-performance bar.
Business impact
Widened ChatGPT's consumer moat, pressured competitor pricing, and opened voice- and vision-driven applications across support, education, and accessibility.
At I/O, Google began rolling out AI Overviews — generated answers above search results — to US users, targeting over a billion users globally by year-end.
Media's fork in the road: News Corp licenses, others litigate
News Corp signed a licensing deal with OpenAI reportedly worth over $250 million across five years — the largest publisher AI deal to date. Weeks later, Forbes and Wired accused Perplexity of republishing paywalled reporting; by July, Perplexity launched a revenue-sharing publisher program.
AI goes on-device: Apple Intelligence and Copilot+ PCs
Apple announced Apple Intelligence at WWDC — on-device AI plus Private Cloud Compute across iPhone, iPad, and Mac, with ChatGPT integrated into Siri. Weeks earlier, Microsoft launched Copilot+ PCs with NPUs for local AI features.
NVIDIA passed Microsoft and Apple to become the world's most valuable public company at roughly $3.34 trillion — having crossed $2 trillion only in February. Shares were up more than 170% on the year.
Claude 3.5 Sonnet and Artifacts: better, faster, cheaper — with a workspace
Claude 3.5 Sonnet beat Anthropic's own larger Opus on many benchmarks at twice the speed and mid-tier price. Artifacts — a side panel where Claude builds and renders code and documents live — debuted alongside it.
Llama 3.1 405B: the first open-weight frontier model
Meta released Llama 3.1, including a 405-billion-parameter model with a 128K context window — the first openly available model widely viewed as competitive with the closed frontier.
Regulation (EU) 2024/1689 took effect as the world's first comprehensive horizontal AI law, with phased deadlines: prohibitions from February 2025, most core obligations from August 2026.
xAI brought its Colossus supercluster online in Memphis — roughly 100,000 NVIDIA H100s assembled in about 122 days — alongside the Grok-2 rollout. In December, xAI closed a $6B round at a ~$45B valuation to keep scaling.
The reasoning era: OpenAI ships o1, then announces o3
OpenAI launched o1, a model that thinks before answering — scoring 83% on an International Math Olympiad qualifier versus GPT-4o's 13%. In December it announced o3, which posted record scores on ARC-AGI and FrontierMath.
Inference-time reasoning emerged as a new scaling axis distinct from model size — capability could now be bought with thinking time.
Business impact
Opened a new competitive front around reasoning, reshaped how enterprises approached complex analytical work, and reset frontier expectations heading into 2025.
At Dreamforce, Salesforce introduced Agentforce — a platform for autonomous AI agents across service, sales, marketing, and commerce. Customers built more than 10,000 agents during the event.
California passes — then vetoes — frontier AI safety bill SB 1047
Governor Newsom vetoed SB 1047, which would have imposed safety testing and shutdown requirements on the largest frontier models, arguing it regulated by size rather than risk. The bill had split the industry — Anthropic cautiously supportive, OpenAI and Meta opposed.
OpenAI closed one of the largest private rounds ever, led by Thrive with Microsoft and NVIDIA participating — with terms reportedly requiring conversion to a for-profit structure within two years.
The Physics prize went to Hopfield and Hinton for foundational neural-network work; Chemistry went to Hassabis and Jumper (with David Baker) for AlphaFold's protein-structure prediction, by then used by over two million researchers.
Amazon committed an additional $4 billion to Anthropic — $8 billion total — with AWS named primary cloud and training partner and flagship models moving to Trainium silicon.
DeepSeek-V3: frontier-class performance at a fraction of the cost
DeepSeek released V3, a mixture-of-experts model trained for a disclosed ~$5.6M in GPU time on export-compliant H800s, performing competitively with leading closed models. The cost figure — later contested as excluding R&D — stunned the industry.
The year of agents. The DeepSeek shock, trillion-dollar compute deals, and AI's first documented impact on employment.
· US · Infrastructure & chips · Funding & markets
Stargate: a $500B bet on American AI infrastructure
At a White House event, OpenAI, SoftBank, Oracle, and MGX announced Stargate — a plan to invest $500 billion in US AI data centers over four years, with $100 billion deploying immediately.
The largest infrastructure commitment in AI history. Frontier AI officially became national industrial policy.
Business impact
Kicked off 2025's data-center construction boom, transformed Oracle into a major AI cloud, and set the template for the year's gigawatt-scale compute deals.
OpenAI released Operator, an agent that controls its own browser to complete tasks — shopping, bookings, reservations — launched on the $200/month Pro tier.
The DeepSeek shock: R1 wipes ~$600B off NVIDIA in a day
DeepSeek open-sourced its R1 reasoning model under MIT license, matching OpenAI o1-level performance at a fraction of the claimed cost. NVIDIA fell 17% — nearly $600 billion, the largest single-day market-cap loss in US history.
Frontier-level reasoning, cheap and open, directly challenged the assumptions underneath hundreds of billions of dollars in US AI capex.
Business impact
Inference prices fell across the industry, open-weight models gained enterprise legitimacy overnight, and US–China AI competition became the market's central theme.
xAI released Grok 3, trained with roughly 10x Grok 2's compute on Colossus's ~200,000 GPUs, claiming benchmark wins over GPT-4o. Grok 4 followed in July.
Claude 3.7 Sonnet and Claude Code: hybrid reasoning meets the terminal
Anthropic shipped Claude 3.7 Sonnet — the first hybrid model offering both instant answers and visible extended thinking — alongside Claude Code, an agentic coding tool that lives in the developer's terminal.
Gemini 2.5 Pro took the lead on math and science benchmarks at launch. At I/O, Google shipped Veo 3 — the first mainstream video generator with native audio: dialogue, sound effects, ambience.
After two years trailing OpenAI, Google led on both reasoning quality and generative media in the same season.
Business impact
Gemini became a serious enterprise API competitor on price-performance, and Veo 3 set off a wave of AI-generated video in advertising and social media.
MCP becomes the industry standard for connecting AI to tools
Anthropic's Model Context Protocol — open-sourced in November 2024 — was adopted by OpenAI in March, Google DeepMind in April, and Microsoft/GitHub in May. By December it had been donated to the Linux Foundation with ~97M monthly SDK downloads.
OpenAI raises $40B at $300B — the largest private round ever
OpenAI closed a $40 billion round led by SoftBank at a $300 billion valuation, disclosing 500 million weekly ChatGPT users. By October, an employee share sale valued it at $500 billion — the world's most valuable private company.
Claude 4: long-horizon autonomy becomes the competitive axis
Anthropic released Claude Opus 4 and Sonnet 4, built for coding and agent workflows, with Opus 4 billed as the world's best coding model and capable of multi-hour autonomous tasks. Sonnet 4.5 followed in September, working autonomously up to 30 hours.
Meta pays $14.3B for Scale AI and starts the talent war
After Llama 4's lukewarm April reception, Zuckerberg bought 49% of Scale AI for $14.3 billion, made Alexandr Wang Meta's first Chief AI Officer, and launched Meta Superintelligence Labs — poaching researchers with packages reported above $100 million.
Vibe coding goes mainstream; the Windsurf saga rewrites AI M&A
Karpathy's term 'vibe coding' became Collins' Word of the Year as Cursor-maker Anysphere hit a $9.9B valuation past $500M ARR. OpenAI's $3B bid for Windsurf collapsed; Google paid $2.4B to license its tech and hire its CEO; Cognition acquired the rest within 72 hours.
Anysphere · Windsurf · OpenAI · Google · Cognition
· China · Model release
China's open-weight wave: Qwen 3, Kimi K2, GLM-4.5, DeepSeek V3.1
Chinese labs shipped a rapid succession of near-frontier open models: Alibaba's Qwen3, Moonshot's trillion-parameter Kimi K2, Zhipu's GLM-4.5, and DeepSeek V3.1. Qwen passed Llama in cumulative downloads by October.
Alibaba · Moonshot AI · Zhipu AI · DeepSeek · MiniMax
· US · Regulation & policy · Infrastructure & chips
Washington's AI Action Plan — and the H20 chip whiplash
The White House published 'Winning the Race: America's AI Action Plan' — 90+ actions across innovation, infrastructure, and AI diplomacy. It followed a year of chip-policy whiplash: H20 sales to China halted in April (a $4.5B charge for NVIDIA), then re-licensed in July for 15% of China revenue.
GPT-5 becomes the default for ChatGPT's hundreds of millions
OpenAI released GPT-5 — its first unified model merging the GPT series with o-series reasoning — as the default for all ChatGPT users, free tier included.
MIT: 95% of enterprise GenAI pilots show no P&L impact
MIT NANDA's 'GenAI Divide' report found that despite $30–40 billion in enterprise spending, ~95% of generative AI pilots delivered no measurable P&L impact — attributing failure to an integration learning gap, not model quality. Purchased solutions succeeded about twice as often as internal builds.
Anthropic triples its valuation to $183B in six months
Anthropic closed a $13B Series F at $183 billion — triple March's $61.5 billion — disclosing run-rate growth from ~$1B in January to over $5B by August, with 300,000+ business customers.
The circular economy of AI: NVIDIA–OpenAI $100B, Oracle $300B, AMD warrants
NVIDIA agreed to invest up to $100B as OpenAI deploys 10GW of its systems; OpenAI confirmed a $300B five-year Oracle compute deal and an AMD agreement with warrants for ~10% of the chipmaker. Total 2025 commitments approached $1 trillion and ~26GW.
Vendors investing in their biggest customer raised circularity concerns at the heart of the boom's accounting.
Business impact
Added hundreds of billions in market value to chip and cloud stocks within days, while analysts flagged circular deals as the systemic-risk indicator to watch.
OpenAI released Sora 2 — video with synchronized audio — inside an invite-only social app built on AI-generated feeds and 'cameos' of real people. It hit #1 on the App Store within 48 hours.
The AI browser war: ChatGPT Atlas and Perplexity Comet
OpenAI launched ChatGPT Atlas, a browser with ChatGPT on every page and an agent mode that completes tasks. Perplexity's Comet, free worldwide from October, had staked the same claim months earlier.
The browser — the gateway to the web's ad economy — became the new battleground for AI distribution.
Business impact
Google accelerated Gemini-in-Chrome, publishers confronted agent traffic that bypasses ads, and enterprises faced new data-security questions about agentic browsing.
OpenAI completes its for-profit restructuring; Microsoft takes 27%
OpenAI recapitalized into a public benefit corporation controlled by its nonprofit foundation. Microsoft received ~27% (worth ~$135B), retained IP access through 2032, and gave up exclusive cloud rights.
AI reaches the workforce: Amazon cuts 14,000 as the evidence mounts
Amazon announced 14,000 corporate cuts — its largest — citing AI investment. Salesforce's Benioff disclosed AI agents let him cut support from 9,000 to ~5,000 roles. A Stanford study found a 13% relative employment decline for workers aged 22–25 in AI-exposed occupations since late 2022.
Ten months after the DeepSeek crash erased $600 billion, NVIDIA crossed $5 trillion — the first company ever — adding roughly a trillion dollars of value in about 100 days.
The year AI became infrastructure. The ROI reckoning and layoff wave continued — but the deeper shift was structural: the model layer commoditized behind standard interfaces, agentic payment rails, the first frontier-lab IPOs, and new inference architectures.
· US · Infrastructure & chips · Company move
OpenAI signs a $10B inference deal with Cerebras
OpenAI announced a partnership reportedly worth over $10 billion under which Cerebras will deliver 750 megawatts of low-latency inference compute through 2028 — wafer-scale chips, not GPUs.
· Global · Research & science · Enterprise adoption
NBER: ~90% of firms report no AI productivity impact yet
An NBER working paper surveying nearly 6,000 senior executives across four countries found roughly nine in ten reported no impact of AI on employment or productivity at their own firms — even though 69% actively use AI.
Oracle cut up to 30,000 jobs — the year's largest single layoff — following Amazon's ~16,000 January cuts. By late May, 2026 tech layoffs reached roughly 142,000, with over half of layoff events explicitly citing AI or automation, as Big Tech redirected payroll toward an estimated $700B combined AI infrastructure buildout.
Qwen nears a billion downloads — and closes its flagship weights
Qwen downloads approached one billion — over half of all open-source model downloads worldwide. Days after open-sourcing Qwen3.6-35B, Alibaba released Qwen3.6-Max-Preview as its first closed-weight flagship, claiming top scores on agentic coding benchmarks.
Chinese open models became the global default — while even open-source champions began holding back frontier weights for commercial advantage.
Business impact
Startups and enterprises standardized on Qwen as base infrastructure, pressuring US lab pricing; the hybrid open/closed strategy became the template to watch.
DeepSeek V4 runs on Huawei silicon: the decoupling milestone
DeepSeek previewed V4-Pro and V4-Flash — open-source, reasoning- and agent-focused, reportedly beating all rival open models on math and coding — and its first models optimized for Huawei's Ascend chips rather than NVIDIA hardware.
GPT-5.5 and GPT-5.5 Pro arrived in the API as agentic models that operate software end-to-end; GPT-5.5 Instant became ChatGPT's default in May, with OpenAI reporting 52.5% fewer hallucinated claims on high-stakes prompts.
New Jersey codifies the ABC test as the federal rule stalls
New Jersey's Department of Labor and Workforce Development adopted regulations codifying how it applies the statutory ABC test to worker classification under the state's Unemployment Compensation Law, Wage and Hour Law and Wage Payment Law. All three prongs must be satisfied: the worker is free from control or direction "both under contract and in fact"; the work is outside the usual course of the business or performed outside all of its places of business; and the worker is "customarily engaged in an independently established trade, occupation, profession or business." The rules become operative October 1, 2026. Acting Labor Commissioner Kevin D. Jarvis said the department "removed provisions in the draft rules that created uncertainty and built a framework shaped by their input."
Classification is where the promise of frictionless global hiring meets the ground. A three-prong conjunctive test decided state by state is not a protocol; it is twenty-odd incompatible dialects with the same name. The compliance floor for anyone paying contractors in the US is set here, in state rulemaking, not in a single federal standard.
Business impact
Contractor-payment and employer-of-record platforms cannot ship one classification logic and call it done — the correct answer depends on the state, and it changes on its own schedule. Prong B in particular is the hard one: if the contractor does the thing the business sells, the arrangement usually fails, no matter how the contract reads.
New Jersey Department of Labor and Workforce Development
· Europe · Regulation & policy · Labor & work
Germany starts drafting as the EU platform work deadline nears
Germany's Federal Ministry of Labour and Social Affairs confirmed it "is currently drafting legislation to transpose the Platform Work Directive into national law," is examining measures on subcontractors engaged by digital labour platforms including the possibility of requiring direct employment, and does not currently plan to extend the directive's scope beyond platform work. Directive (EU) 2024/2831 on improving working conditions in platform work was adopted in October 2024 and applies from December 2, 2026. Rather than fixing criteria centrally, it requires each member state to establish an "effective legal presumption" of employment under national law, and creates what the European Parliament describes as "the first EU rules on the use of artificial intelligence in the world of work," including transparency obligations on algorithmic decisions and a prohibition on processing emotional or psychological data.
The directive was supposed to harmonise platform work across Europe, and in the most important respect it does the opposite: the employment presumption is delegated to twenty-seven national legislatures, each defining its own triggers. A protocol that every implementer configures differently is a standard in name only — and the algorithmic-management provisions arrive on top, reaching further into automated decisions than the AI Act does.
Business impact
Employment and payroll infrastructure operating in Europe faces a compliance surface that fragments rather than consolidates before December 2026, with the burden of proof on classification shifting toward the platform. Building to a single European standard is not an option, because there will not be one.
European Union · German Federal Ministry of Labour and Social Affairs
· Europe · Regulation & policy
EU delays AI Act high-risk rules in the Digital Omnibus deal
The EU institutions agreed the first amendments to the AI Act: high-risk obligations postponed to December 2027 and beyond, while general-purpose AI supervision and enforcement still begin August 2026.
A pragmatic admission that the oversight infrastructure wasn't ready — the world's most ambitious AI law recalibrated under simplification pressure.
Business impact
Thousands of companies gained 16+ months of compliance relief, while frontier model providers still face enforceable obligations and fines from August 2026.
Cerebras IPO: biggest US tech listing since Uber, up 68% on debut
Cerebras priced above range, raised ~$5.55 billion, and jumped 68% on debut to a ~$95 billion valuation — on $510M of 2025 revenue and the OpenAI deal as anchor contract.
Google announced the Gemini 3.5 family — Flash immediately available across Search, the Gemini app, and its developer stack; Pro with Deep Think and a reported 2M-token context to follow — plus Gemini Omni, which takes any modality in and outputs video.
NVIDIA's record quarter: $81.6B revenue, market cap past $5T
NVIDIA reported record revenue of $81.6 billion (up 85% year over year) and $58.3 billion net income, guiding to ~$91 billion for the current quarter — the clearest evidence that 2026's ~$700B hyperscaler capex was still converting into chip revenue.
MCP drops sessions — the agent protocol becomes ordinary plumbing
The Model Context Protocol locked the release candidate for its 2026-07-28 specification, with final publication scheduled for July 28, 2026. The revision makes the protocol core stateless: the initialize/initialized handshake is removed, and "the Mcp-Session-Id header and the protocol-level session that came with it are also removed," so servers can handle requests on any instance rather than requiring sticky routing and shared session stores. It adds a formal Extensions framework with reverse-DNS IDs and two official extensions (MCP Apps, for sandboxed-iframe interfaces, and Tasks), six authorization proposals aligning with OAuth 2.0 and OpenID Connect, full JSON Schema 2020-12, W3C Trace Context propagation, and a lifecycle policy requiring "at least twelve months between deprecation and the earliest possible removal." Roots, Sampling and Logging enter deprecation.
Every one of these changes is boring: load balancing, standard auth, trace propagation, a published deprecation window. That is exactly the point. Protocols stop being demos when they acquire the unglamorous properties that let ordinary operations teams run them — and the value migrates off the model and onto the interfaces and workflows built around it.
Business impact
Removing the stateful handshake is what makes remote MCP servers deployable behind commodity infrastructure, which lowers the cost of putting agents into production for teams without specialist platform engineering. A twelve-month deprecation floor is also a procurement signal: it makes the protocol safe to build a roadmap on.
Anthropic raises $65B at $965B, overtakes OpenAI, files for IPO
Anthropic closed a $65 billion Series H at a $965 billion valuation — surpassing OpenAI's for the first time — with run-rate revenue past $47 billion. Days later it confidentially filed a draft S-1, setting up a potential listing as early as October 2026.
The world's most valuable AI startup, and the first frontier-lab IPO filing — a watershed for public-market AI exposure.
Business impact
Reset private AI valuations near the trillion-dollar mark and pressured OpenAI's own path to public markets, driven by Claude's enterprise coding momentum.
Apple makes the model layer a swappable OS primitive
At WWDC 2026 Apple opened its Foundation Models framework so developers can call any model — Apple's on-device models or cloud models like Claude and Gemini — through a single Swift LanguageModel protocol, switching providers without code changes. Xcode 27 added agentic coding with built-in routing to Anthropic, Google, and OpenAI agents, and Apple's Foundation Models on Private Cloud Compute became free for small App Store developers.
It commoditizes the model layer at the level of the operating system. A provider-neutral interface baked into the platform SDK turns the model into an interchangeable component and makes the surrounding workflow — distribution, integration, trust — the durable asset.
Business impact
Apple's developer base becomes a distribution channel for Claude, Gemini, and OpenAI rather than a closed garden, and free on-device-plus-cloud inference removes the cost barrier for millions of small developers — pressuring every platform to treat models as swappable.
OpenAI confirmed it had confidentially submitted a draft S-1 to the SEC — the first formal step toward an IPO — with Goldman Sachs and Morgan Stanley managing. The company, last valued near $852 billion, said timing was undecided. The filing came roughly a week after Anthropic's own confidential S-1.
Both leading frontier labs entered the public-markets pipeline within two weeks. That turns 'private AI valuations' into an imminent question of public disclosure: audited financials, named risk factors, and the first comparable look at frontier-lab economics.
Business impact
Public-market AI exposure moves from NVIDIA and the hyperscalers toward the model labs themselves, and the disclosure an IPO forces will set the terms of the bubble-or-supercycle debate with numbers instead of narrative.
Claude Fable 5: the first broadly available Mythos-class model
Anthropic released Claude Fable 5 — the first model in a new Mythos-class tier above the Opus line — scoring more than 10% above Opus 4.8 on some software-engineering benchmarks, with broad release enabled by new safeguards in high-risk domains.
Google releases DiffusionGemma, an open text-diffusion LLM
Google released DiffusionGemma, an open-weight (Apache 2.0) text-diffusion model on the Gemma 4 backbone — a ~26B mixture-of-experts (~3.8B active) that denoises 256-token blocks in parallel rather than generating one token at a time. Google reported 1,000+ tokens/sec on a single H100 (up to ~4x faster than comparable autoregressive models) with native vLLM support; published quality trails standard Gemma 4 while latency and infilling improve.
It makes a non-autoregressive generation path real, open, and deployable on mainstream serving infrastructure for the first time — a different lever on inference cost and speed than building a bigger model. Whether the paradigm holds is unproven: quality lags and adoption is unknown.
Business impact
If diffusion decoding proves out, it shifts the latency and cost economics of inference for local and high-throughput serving. For now it is a significant architectural milestone to watch, not a settled breakthrough.
Mastercard launched Agent Pay for Machines, an open protocol letting AI agents and software systems pay one another — including sub-cent micropayments — across cards, bank accounts, and stablecoins, with identity verification, enforced spending limits, and guaranteed settlement. Agent credentials and permissions are stored on public blockchains (Polygon, Solana, Base); roughly 31 launch partners included Coinbase, Stripe, Adyen, and Cloudflare.
The hard problem in agentic commerce is not moving money — it is proving an agent's identity and constraining its authority. Putting that identity, consent, and permission layer on shared infrastructure is the trust standard the agent economy was missing.
Business impact
A card network stakes out the trust-and-settlement layer of machine-to-machine commerce, framing it as long-term infrastructure rather than near-term revenue — and pulling Stripe, Visa, and Google into a race to define how autonomous software transacts.
· US · Regulation & policy · Labor & work · Company move
Anthropic publishes an economic policy framework for AI displacement
Anthropic published an Economic Policy Framework mapping government responses to three escalating AI-unemployment scenarios (5%, 10%, and an 'unprecedented' level) — contemplating wage insurance, retraining, capital accounts, UBI, and equity-sharing, financed by taxes on 'relevant companies.' CEO Dario Amodei's companion essay argued the US government should be able to block release of frontier models that fail independent safety testing. Anthropic committed $350M: a $200M Economic Futures Research Fund and a $150M early-career fellowship.
A market-leading developer is formally proposing to be taxed to fund — and to have its releases gated by — the disruption its own products may cause. Read the incentives: the disruptor is moving first to shape the rules of its own disruption.
Business impact
It reframes the AI-labor debate from 'will jobs go' to 'who pays and how it's governed,' and — paired with new measurement infrastructure like Stanford's AI Economic Indicators — gives the displacement question primary-source anchors from inside the industry.
RWS grows profit 33% as AI reaches a third of revenue
RWS, the largest listed language services company, reported revenue of £360.3m for the six months to 31 March 2026 — up 5% reported and 7% organic constant currency — with adjusted profit before tax up 33% to £24.0m from £18.0m. AI-related products and services rose to 32% of group revenue, from 26% a year earlier. The company launched Language Weaver Pro in partnership with Cohere, citing benchmarking tests showing it outperforming leading AI translation tools, and acquired Obviously Group in early May for £16.5m initial cash consideration, capped at £40m in total.
Profit grew six times faster than revenue while the AI share of the mix climbed — the opposite shape to the compression showing up in customer-experience outsourcing the same month. Two labour-intensive service industries, the same technology, opposite margin directions. Whether AI compresses a service business or expands it is not a property of the technology; it turns on whether the firm bills for the output or for the hours.
Business impact
Gives service-business operators a live counter-example to the assumption that AI necessarily deflates language and support pricing, and a concrete model: build the AI into a product line that carries its own margin rather than using it to discount the existing one.
SpaceX (xAI inside) prices the largest IPO on record
SpaceX priced the largest IPO on record — roughly $75 billion raised (555.6 million shares at $135), valuing the company near $1.8 trillion — and closed its Nasdaq debut up ~19% under the ticker SPCX. The listed entity includes xAI, which merged into SpaceX in an all-stock deal in February 2026, bringing the Grok models and the Colossus data center (220,000+ NVIDIA GPUs) onto the public company's balance sheet.
Read through the AI lens, not the rocket: public-market investors now get frontier-AI compute exposure (xAI, Colossus) bundled inside the largest listing in history — the capstone of an AI capital cycle going public in a single quarter (Cerebras → Anthropic → OpenAI → SpaceX/xAI).
Business impact
One of the world's largest AI training clusters becomes a public asset, and the scale of the raise resets the ceiling for AI-adjacent listings — concentrating still more market exposure in the AI infrastructure trade.
Salesforce buys Fin for $3.6B on a 76% resolution claim
Salesforce signed a definitive agreement to acquire Fin — the AI customer service company that renamed itself from Intercom — for approximately $3.6 billion, subject to customary purchase price adjustments, folding it into Agentforce. Salesforce's release states Fin resolves "on average 76% of support volume end-to-end" for more than 30,000 companies, running on Apex, Fin's proprietary model purpose-built for customer support. The deal is expected to close in the fourth quarter of Salesforce's fiscal year 2027.
The largest consolidation yet in AI customer support, and the number at the centre of it deserves reading carefully. A resolution rate measures volume that never reached a human. It does not measure whether the customer's problem was solved, whether they came back three days later, or whether they simply gave up. Those are different questions, and only the first one is easy to instrument — which is precisely why it is the one that gets quoted.
Business impact
Puts a public benchmark into every support procurement conversation: buyers will now be asked why their own deflection rate is below 76%, and vendors will be measured against a figure whose definition is set by the seller. Teams that cannot separate contained volume from resolved problems will negotiate against a number they cannot audit.
· Global · Regulation & policy · Research & science
IMF: Nigeria is 60% of sub-Saharan Africa's stablecoin inflows
The IMF published Stablecoins in Nigeria, finding that Nigeria has accounted for roughly 60% of stablecoin inflows in sub-Saharan Africa since 2019. The report describes the shift plainly: "What began as a niche technology has become a meaningful cross-border payments channel. Its rapid growth is easing long-standing frictions in cross-border transactions." Central Banking reported the IMF set out four priorities for managing the risks, summarising its position as allowing stablecoin innovation while managing the exposure it creates.
This is a multilateral institution, not a vendor, reporting that dollar-denominated stablecoins have become a real payments channel — and the reason is not enthusiasm for the technology. It is that correspondent banking works badly in exactly these corridors. Adoption is highest where the incumbent rails are worst, which is the honest version of the story.
Business impact
For anyone paying contractors or staff into emerging markets, this reframes stablecoin rails from speculative to plausible — while naming the cost: a payout channel denominated in dollars is a small act of dollarisation, and the receiving country's central bank has a legitimate objection to it. The rail solves your problem and creates someone else's.
DeepL buys Mixhalo, moving translation toward live interpretation
DeepL acquired Mixhalo, a real-time audio platform for live events that had raised over $39 million from investors including Fortress Investment, Founders Fund, Defy Partners and Cowboy Ventures. Terms were not disclosed. CEO Jarek Kutylowski said the platform "will allow us to show how DeepL's tech works in real time and in environments like conferences where people are present on the ground." DeepL launched voice-to-text translation in over 33 languages in 2024 and a voice-to-voice translation suite in April 2026, and is opening a San Francisco office alongside the deal.
Machine translation has been encroaching on written work for a decade. Live simultaneous interpretation has held out, because it runs in real time, under social pressure, with no edit pass and no second attempt — the conditions under which tacit skill matters most. Buying a live-event audio platform is a statement about which direction the frontier is moving.
Business impact
Interpretation providers should expect the low-stakes end of live work — conference sessions, all-hands, webinars — to come under price pressure first, while regulated, high-consequence settings such as medical and legal interpreting stay human for longer. The defensible position is the work where being wrong is expensive.
The OCC proposes the compliance perimeter for stablecoin issuers
The OCC issued Bulletin 2026-28, a notice of proposed rulemaking titled "GENIUS Act: Anti-Money Laundering/Countering the Financing of Terrorism and Sanctions Compliance." It would require OCC-supervised permitted payment stablecoin issuers "to comply with the BSA, sections 4(a)(5) and 4(a)(6)(B) of the GENIUS Act, and applicable regulations issued by the Financial Crimes Enforcement Network (FinCEN) and the Office of Foreign Assets Control, including any anti-money laundering and countering the financing of terrorism (AML/CFT) program, sanctions program, and reporting requirements." It applies to federal qualified issuers and to state qualified issuers over which the OCC has authority, creates a supervision and enforcement framework, and establishes a consultation process between the OCC and FinCEN before significant AML/CFT actions. Comments run 30 days after Federal Register publication.
This is the boring half of the stablecoin story and the half that decides whether regulated businesses can actually use these rails. Nothing about a payout channel matters operationally until an issuer sits inside a supervised AML, sanctions and reporting regime. Compliance is not a tax on the product here — it is the feature that makes the product usable by anyone with a real balance sheet.
Business impact
For payroll and cross-border payout infrastructure, this is the difference between a stablecoin rail being an experiment and being procurable: the counterparty becomes a supervised entity with a named regulator and an enforcement path. It also sets the diligence bar — issuer status under the GENIUS Act becomes a question every serious buyer asks.
Office of the Comptroller of the Currency · FinCEN
· US · Funding & markets · Enterprise adoption
Concentrix cuts guidance while AI deals grow 400%
Concentrix reported fiscal Q2 2026 revenue of $2,462.5 million, up 1.9% as reported and 0.6% in constant currency, with operating cash flow of $257.9 million. It lowered full-year constant-currency revenue growth guidance to 0.25%–1.25% and non-GAAP operating income guidance to $1,200–1,230 million. The same release reported "iX Suite deals up 400% year over year," with CEO commentary that "Our blended AI and services approach is delivering value to clients by lowering their costs and increasing their revenue, helping us differentiate ourselves in the marketplace."
Read the two halves of the release against each other. The AI product line is growing fast in percentage terms off a small base; the services business still sets the guidance, and the guidance came down. This is what the middle of a service-to-software transition looks like on an actual income statement — not a clean pivot, but a fast-compounding new line that cannot yet carry the company, disclosed in the same document as a cut.
Business impact
For operators, it prices the transition risk honestly: helping clients lower their costs is a good sentence in a press release and a difficult one in a revenue model when you are paid for the work being removed. The lesson is to reprice toward outcomes before the volume repricing arrives, not after.
Visa, M-PESA and Onafriq pilot stablecoin settlement in the DRC
Visa, M-PESA Africa and Onafriq began piloting the use of stablecoins for cross-border mobile money transactions in the Democratic Republic of the Congo. The stated aim is to use blockchain-backed digital currency to make cross-border settlement faster and cheaper, addressing what the reporting describes as the "high transaction costs, lengthy processing times and limited accessibility" of traditional banking corridors. The DRC was selected as the initial testing ground; this is a pilot, not general availability.
The interesting party here is Visa. When a card network and a mobile money operator run settlement over stablecoins, the technology has stopped being a challenger to the incumbent rails and started being a component inside them. That is usually how infrastructure actually changes — not by replacement, but by quiet substitution one layer down, while the interface the user sees stays the same.
Business impact
If settlement over stablecoins works at mobile-money scale in a thin-banking market, the cost and speed floor for paying people in similar corridors moves. The constraint on global payouts has never been willingness to pay; it has been the correspondent chain in between.
Premium publisher ad supply falls up to 41% as AI search bites
Benchmarking data from Ozone covering roughly 20 billion impressions across premium publishers including the Guardian, News UK and Dow Jones' Wall Street Journal showed Q2 2026 ad request volumes down roughly 32%–37% year over year in the US and 39%–41% in the UK. Combined US and UK programmatic spend fell 30.6% year over year in the first half of 2026 — US down 44%, UK down 14.3% — while June average eCPMs ran about 30% higher year over year in the UK and about 7% higher in the US.
An ad request is a count of pages actually served, not a modelled estimate of clicks foregone, which makes this one of the cleaner available measurements of what AI-mediated search is doing to the open web. Prices rising on collapsing volume is a yield-led market: publishers are extracting more per impression precisely because there are fewer of them, which is not a durable position.
Business impact
For anyone whose reach depends on being discovered through search, the intermediary layer is thinning. The practical response is to own the destination — an archive, a list, a body of work under your own domain that can be cited directly — rather than renting attention from a referral channel that is being disintermediated.
Google AI Mode starts acting inside third-party apps
Google began rolling out the ability for AI Mode to link to and act inside select third-party apps, starting with Instacart, Canva and YouTube. Users can add recipe ingredients "directly to your shopping cart and quickly check out on the Instacart app or website," request design templates from Canva, and curate a playlist and "instantly save it to YouTube Music." The update is rolling out to users in the US.
Search has spent twenty-five years as a referral layer: it told you where to go and you went there. Acting inside the destination collapses that step. The visit — the thing publishers, retailers and every content business monetise — becomes optional, and the surface that captures the intent is no longer the one that fulfils it.
Business impact
It compounds the referral decline showing up in publisher inventory: fewer sessions reach the destination at all, and the ones that do arrive further down the funnel. Businesses that treated search as their distribution now have to be the destination people ask for by name.
The same four questions for every industry: what changed, what drove it, what businesses now do differently, and what it means if you are building something.
Software development＋−
What changed
The most transformed profession of the era. Code completion gave way to agentic tools that plan, write, test, and ship multi-hour units of work. By 2025, AI wrote a large share of new code at major companies, Stack Overflow traffic had collapsed, and entry-level engineering hiring contracted measurably.
What drove it
GitHub Copilot, ChatGPT, Claude (especially Claude Code and the Claude 3.5+ line), Cursor, Devin, OpenAI Codex, and the MCP standard that let agents reach real tools.
What businesses do differently
Teams budget for AI tooling per engineer the way they once budgeted for cloud. Senior engineers spend more time reviewing and specifying, less time typing. 'Vibe coding' made prototyping nearly free, so the bottleneck moved to judgment: what to build and what to trust.
For founders and leaders
Software is cheaper to produce and easier to copy than ever. Velocity is table stakes; durable advantage comes from distribution, data, and taste. If your engineers aren't using agentic tools daily, you are paying a competitor's payroll.
The first business function with documented, at-scale labor substitution. Klarna's assistant doing the work of 700 agents (2024) and Salesforce cutting support from 9,000 to ~5,000 roles (2025) made the new economics undeniable: AI handles the routine majority, humans handle exceptions and high-value moments.
What drove it
GPT-4-class models via API, Salesforce Agentforce, purpose-built support platforms (Intercom Fin, Decagon, Sierra), and the Brynjolfsson field study showing novice agents gained the most from AI assistance.
What businesses do differently
Support became a deflection-rate and resolution-quality discipline rather than a headcount discipline. Companies route by complexity, measure AI containment honestly, and rebuild knowledge bases as machine-readable sources of truth.
For founders and leaders
Support economics changed permanently — plan for AI-first support from day one, with human escalation as a designed experience rather than a fallback. For BPO and CX businesses, the move up the value chain (quality, judgment, integration) is existential.
Content production costs collapsed — copy, imagery, and by 2025 full video ads with sound (Veo 3, Sora 2). Meanwhile AI search (Overviews, ChatGPT, Perplexity) began replacing clicks with answers, breaking the SEO-to-traffic pipeline that funded two decades of content marketing.
What drove it
ChatGPT and GPT-4o for copy, Midjourney and Adobe Firefly for imagery, Veo 3 and Sora for video, Google AI Overviews and AI browsers for discovery, plus AI SDRs and CRM agents (Einstein GPT, Agentforce).
What businesses do differently
Brands optimize to be cited by AI engines, not just ranked by Google. Creative teams shifted from production to direction and judgment. Outbound sales went AI-personalized at scale — which devalued mass personalization almost immediately.
For founders and leaders
Assume your buyers ask an AI before they ever reach your site. Being the credible, citable source in your category now matters more than volume publishing. Distribution moats built on SEO arbitrage are melting.
AI went from back-office experiment to front-office tool: Morgan Stanley's GPT-4 advisor assistant and BloombergGPT arrived within months of ChatGPT. By 2025–26, AI also became the market itself — NVIDIA's rise to $5T, trillion-dollar circular compute deals, and the AI capex cycle dominating equity markets.
What drove it
GPT-4-class models with retrieval over proprietary research, BloombergGPT, reasoning models for analysis, and agent platforms for operations like reconciliation and compliance checks.
What businesses do differently
Firms deploy AI assistants over internal knowledge with strict audit trails; analysts draft with AI and verify by hand. CFOs now evaluate AI spending like capex — and investors price 'AI exposure' as a factor in every portfolio.
For founders and leaders
Regulated-industry AI rewards whoever solves trust and auditability, not just capability. And if you raise capital, understand that the AI cycle now sets the market's temperature — plan around its volatility.
AI crossed from administrative tooling into science itself: AlphaFold 3 modeled drug-molecule interactions, its creators won the 2024 Nobel in Chemistry, and ambient clinical documentation became the fastest-adopted hospital software in memory.
What drove it
AlphaFold and Isomorphic Labs in drug discovery, ambient scribes (Nuance DAX, Abridge), GPT-4-class models for clinical documentation, and reasoning models cutting hallucinations on medical questions.
What businesses do differently
Health systems buy AI to give clinicians time back — documentation first, diagnosis support carefully and slowly. Pharma R&D treats computational structure prediction as standard infrastructure.
For founders and leaders
The wedge is workflow relief, not diagnosis. Sell time back to clinicians, prove safety, and treat regulatory rigor as your moat — healthcare punishes shortcuts harder than any other market.
ChatGPT reached students faster than any institution could respond — first as a cheating panic, then as curriculum. Tutoring economics inverted: a personal, always-available tutor became effectively free, while assessment design had to assume AI assistance everywhere.
What drove it
ChatGPT (free GPT-4o mattered most), Claude for long-document study, Khanmigo and AI tutoring products, and voice-mode models that made conversational learning natural.
What businesses do differently
Institutions moved from banning to integrating: AI-resistant assessment (oral, in-person, process-based), AI literacy as a subject, and personalized practice at scale. Corporate training quietly became one of AI's strongest use cases.
For founders and leaders
The opportunity is not 'ChatGPT for school' — it is rebuilding assessment, credentialing, and skill verification for a world where output is no longer proof of understanding.
The cost of producing images, video, and audio fell toward zero — photoreal images in 2023, coherent video in 2024, video with native sound in 2025, and a TikTok-style feed of pure AI content (Sora app) by late 2025. Meanwhile publishers split between suing AI companies and licensing to them.
What drove it
Midjourney, Adobe Firefly, Sora and Sora 2, Veo 3, Runway, ElevenLabs for voice — plus the News Corp–OpenAI licensing template and the SAG-AFTRA consent-and-compensation framework.
What businesses do differently
Studios and agencies restructured around AI-assisted pipelines with human creative direction. Newsrooms negotiated licensing while fighting traffic loss from AI answers. Provenance and verification became product features.
For founders and leaders
When production is free, curation, taste, and trust become the business. Owning a relationship with an audience — not a content library — is the asset that survives.
GPT-4 passing the bar exam in early 2023 made law the symbol of white-collar exposure. Contract review, research, and drafting compressed from days to hours — while early hallucinated-citation scandals taught the profession exactly where the limits were.
What drove it
GPT-4 and Claude's long-context document analysis (hundred-page contracts in one prompt), Harvey and CoCounsel built on frontier APIs, and retrieval systems over firm precedent.
What businesses do differently
Firms adopted AI for first drafts and review with verification workflows; clients began questioning billable hours for work AI accelerates. In-house teams insource more routine work.
For founders and leaders
Legal AI rewards domain-deep products with verification built in. For everyone else: AI-accelerated legal review is cutting transaction friction — contracts, compliance, fundraising — faster than most founders realize.
AI moved from a technology decision to the central strategy question. Boards demanded AI plans in 2023; by 2025–26, leaders were publicly tying headcount, capex, and org design to AI — while studies (MIT's 95%, NBER's 90%) showed most deployments still hadn't reached the P&L.
What drove it
Copilots embedded in productivity suites, enterprise agent platforms (Agentforce, ServiceNow), MCP-connected internal tools, and the public example set by AI-forward operators like Klarna, Salesforce, and Amazon.
What businesses do differently
Leaders flattened org charts, slowed entry-level hiring, and redirected payroll into AI infrastructure. The winning pattern from the research: buy focused solutions, embed them in one workflow, measure honestly — rather than broad internal platform builds.
For founders and leaders
The gap between AI adoption and AI impact is where competitive advantage lives right now. Pick narrow workflows, instrument them, and compound. Announcing an AI strategy is not having one.
The leverage available to a tiny team exploded. Tasks that required hires or agencies — copywriting, design, bookkeeping prep, support, basic software — became AI-assisted or AI-done. Stripe's data showed top AI startups reaching $1M revenue in a median 11.5 months, faster than the best SaaS cohort ever.
What drove it
ChatGPT as the universal assistant, open models (Llama, Qwen, Mistral) keeping costs near zero, vibe-coding tools turning ideas into working software, and agent products automating admin.
What businesses do differently
Solo founders run functions that used to be departments. 'How big does this team really need to be?' became the default planning question. The tiny, AI-leveraged company became a legitimate ambition rather than a constraint.
For founders and leaders
Your effective headcount is your team times the leverage of its tools. Competitors have the same leverage — so speed of learning, not access to AI, is the differentiator that remains.