The State of AI in 2026: Who’s Winning, Who’s Losing, and Why Gen Z Is Checking Out

Introduction

Three years ago, if you had picked a random person at a tech conference and asked them what 2026 would look like, the answer would almost certainly have involved flying cars, AI that writes its own code, and a world run partly by Large Language Models. None of those things have quite materialised as predicted — though one of them is closer than you think. The actual picture of AI in mid-2026 is messier, more contradictory, and in some respects far more interesting than the clear-eyed "AI is eating the world" narrative Silicon Valley has been selling since late 2023.

This article is a grounded, non-political roundup of what is actually happening across the technology spaces that define our near future: the AI model wars and cloud economics, the data centre boom nobody expected at this scale, the automation tools quietly reshape how design and engineering teams work, the security cost of AI-era software practices, and the unexpected demographic backlash one Silicon Valley legend probably did not see coming.

1. AI Model Wars: Tencent, Alibaba, DeepSeek, and the Open Source Surge

The New Rivals Are Not From California

For the first half of the generative AI era, the conversation was dominated by San Francisco Bay Area firms: OpenAI's ChatGPT, Google's Gemini, Anthropic's Claude, and the smaller players around Meta's LLaMA. That dynamic started to fracture in 2025 and came into full view in 2026. Chinese AI companies Tencent and Alibaba have emerged as heavyweight model builders at scale — with Alibaba's Qwen 3 stacked up alongside, and occasionally ahead of, Western counterparts in both reasoning benchmarks and coding evaluations. Tencent's Hunyuan series is closing that gap rapidly.

DeepSeek — a Chinese research and product company that started making noise in late 2025 — produced benchmark results in early 2026 that forced many US-based AI observers to revisit the "-China-is-falling-behind" narrative that dominated most of 2024. DeepSeek's R2 inference model demonstrated coding and math reasoning performance on a level that, on paper, rivalled Claude Opus at a fraction of the compute cost. The significance was not simply that another company had built a competitive model, but that doing so required notably less expensive infrastructure. That has enormous implications for global compute economics.

Open Source Is Finally Eating Into Proprietary Models

The open-source AI ecosystem — largely driven by Meta's LLaMA series and the vibrant independent community around Hugging Face — has crossed a critical competency threshold. LLaMA 4-class fine-tunes, made available fully permissively or at low cost via platforms like Together.ai and Fireworks.ai, are now capable enough that many downstream developers are choosing community-hosted models rather than paying for GPT-4.5 or Claude Opus APIs. For use cases where fine-grained cost control, offline deployment, or regulatory requirements around data sovereignty matter, this is a decisive shift. Enterprise procurement teams, historically slow-moving, began documenting open-source-first AI policies in Q1 2026, and several Fortune 500 engineering orgs have now migrated at least 30% of their production inference workloads to LLaMA-based stack.

Open-source language models have also driven a corresponding wave of innovation in open-weight image and audio models. Stability AI, after two years of financial turbulence, stabilised its leadership and shipped Stable Diffusion 4 in public beta, which many professional photographers now use as a reference generator or workflow tool. The open-weight audio speech synthesis market meanwhile moved toward parity with ElevenLabs with rapid open alternatives such as the Coqui TTS series and Microsoft's VALL-E 2 open checkpoint batches.

The Silicon Valley Bank of Compute — Nvidia

The model war is backed by a compute war, and Nvidia is essentially the only bank that matters there. In Q1 fiscal 2027 (ended April 2026), Nvidia reported record overall revenue of $81.6 billion and record data centre revenue of $75.2 billion — a year-over-year jump of 92%. The scale of those numbers is difficult to contextualise: it exceeds the annual revenue of most entire industrial sectors. That revenue is almost entirely tied to AI training and inference workloads. Nvidia's Blackwell B200 and GB200 architecture chips have become the de-facto standard hardware for both hyperscaler and enterprise training clusters. Every major AI lab and cloud provider is building around Nvidia's product roadmap, creating both massive upside for Nvidia and concentration risk for the broader AI infrastructure ecosystem.

Nvidia's dominance is so pronounced that it has begun to concern antitrust regulators on two continents. The US Federal Trade Commission issued a formal information request to Nvidia in late 2025 around competitive practices in the AI GPU market. Meanwhile, the European Union's Digital Markets Act is now being applied to the CUDA software ecosystem. Despite the regulatory headwinds, Nvidia's Q1 2027 numbers suggest that demand — particularly from hyperscalers building foundation models and enterprise customers orchestrating inference at scale — is running well ahead of supply.

TechCrunch reported in late May 2026 that the company drove record overall revenue of $81.6 billion in its fiscal Q1 2027, with data centre alone generating $75.2 billion, representing a 92% increase from the prior year. That number alone says more about the state of AI economics in mid-2026 than any single product launch.

2. Data Centres: The Silent Infrastructure Revolution

For three years, "data centre expansion" was a phrase most readers glossed over. In 2026, it is the single most consequential — and least discussed — infrastructure trend in technology. AI training runs, particularly for large foundation models, consume megawatts of power. Inference at the user layer — where billions of daily ChatGPT, Gemini, and Claude interactions happen — is adding a permanent, growing energy footprint to every cloud provider's bill.

Feeling the heat locally, several US states and municipalities have started pushing back against new data centre construction. The Verge reported in May 2026 that AI data centre projects are continuing to appear across the United States, frequently encountering opposition from communities concerned about local grid strain, water consumption (training clusters draw enormous amounts of cooling water), and noise from emergency generators. Meanwhile, the infrastructure buildout continues. The Wall Street Journal, running its own AI quiz for readers to gauge general data centre literacy, captured an unspoken reality: data centre economics are now a mainstream topic, even if most people cannot name a single hyperscaler company besides Amazon or Google.

The economics of the data centre arms race are also quietly transforming the electricity grid. Vistra Corp, Constellation Energy, and NextEra Energy — previously treated as stodgy utility operators — are now the critical political and economic intermediaries for AI. The Verge reported in May 2026 that there were multiple competing proposals related to NextEra's potential acquisition of Dominion Energy, a deal that would fundamentally reshape how the eastern US grid interacts with data centre construction. Globally, the International Energy Agency issued an updated forecast in early 2026 suggesting that data centre electricity consumption could double by 2030, with AI as the primary growth driver.

The intersection of AI compute, energy, and data centre economics is now sufficiently entangled that a full financial analysis of any AI company today has to start with the energy market, not just software economics. Investors who focused purely on model quality through 2024 have had to rapidly adjust their mental models to account for compute costs, data centre supply chains, and energy commodity exposure.

3. AI Agents: From Research Paper to Product

The Agent Wave Is Not Hype — It Is Just Getting Started

While the big language model players continued to iterate on base model quality, 2026 became the year AI agents moved from research paper novelty to a real mainstream product category. The most significant signal arrived in May 2026 when Figma announced its native AI agent — directly available inside the design canvas, alongside the tools that designers already use every day.

The Figma agent is not a side feature bolted on top of the platform. It is a purpose-built agentic layer tied to design context in ways that are not possible with external or third-party AI integrations. The agent understands your components, your design tokens, your library, and your team's conventions, and it can execute bulk-edits, iterate on variants, darken every frame to check contrast, rename variables consistently, generate placeholder content at scale, and apply stakeholder feedback across multiple frames and screens simultaneously. The team calls the framing "false choices are emerging: speed or precision? AI generation or direct manipulation? You should not have to choose." And the intent is genuine — the agent sits inside the same file and workspace as the team, meaning the context lives there natively without handoff overhead.

This matters enormously for the future of design tooling. For years, the tension in design software has been between the craftsmanship of direct manipulation — clicking layers, dragging, nudging — and the promise of generative AI that could eliminate tedious rework. The first round of AI integrations usually failed simply because they lacked the context that a human designer already holds. By embedding the agent inside the Figma canvas, directly connected to the team's workspace, Figma has effectively closed the handoff gap. Adobe, Canva, and several other design tools racing to catch up built their AI layers on top of external APIs. Figma's native agent architecture is likely to prove a durable differentiation.

MCP: The Protocol That Ties Everything Together

Behind the Figma agent announcement sits a quieter but perhaps more foundational development: the Model Context Protocol (MCP), an open specification that allows AI agents to read and write structured data from real tools and data sources without brittle, hand-built integrations. Developed by Anthropic and released as an open standard with a growing ecosystem of implementations, MCP is quietly becoming the HTTP of the AI agent world. Figma's MCP server allows the agent to pull code back into Figma from a codebase, iterate on it using the canvas, and push it back — so prototyping and implementation can happen in a tight loop, assisted by AI but controlled by the team.

Design and professional creative tools are not the only category where agentic AI is landing. GitHub Copilot, already in widespread use, moved to embed multi-turn planning and codebase-wide architectural awareness — rather than single-function completions — in its 2026 spring update. AWS Bedrock, Google Vertex AI, and Azure AI Studio all shipped agent orchestration layers in Q1 2026. The pattern is clear: AI is moving from a Copilot metaphor — which implies a second driver sitting beside you — toward an Agent metaphor, in which the AI takes on bounded, accountable responsibility for a task.

4. Generative AI in Everyday Workplaces

The Intuit Layoff as a Marker of AI-Enabled Autom

In May 2026, Intuit — the $60 billion financial software company behind TurboTax and QuickBooks — announced it would cut approximately 3,000 jobs, roughly 17% of its global workforce. CEO Sasan Goodarzi framed the cuts explicitly in terms of AI investment: the company would redirect capital toward embedding AI across its product lines. This reading of careful what is actually happening: Intuit is not shrinking. It is investing more heavily in AI, and the infrastructure that underpins those AI investments — particularly structured automation of routine financial tasks — is replacing jobs faster than it is growing headcount elsewhere.

Intuit's move is a data-point in a landscape-wide pattern that economists have been watching since late 2024. Professional services firms across law, accounting, insurance underwriting, and back-office operations are systematically re-engineering repetitive work around AI systems. In law, Harvey AI expanded its practice-wide adoption in Am Law 200 firms. In accounting, automated reconciliation stacks running on Anthropic and OpenAI API have reduced month-end close cycles at multiple mid-market firms. The pattern is consistent: where work is structured, repetitive, and involves processing or summarising large volumes of text or data, AI is quietly creeping in.

What makes the Intuit moment distinct is the explicit framing: "This is not about replacing people — it is about investing in AI to serve customers better." That phrasing — deployed precisely in May 2026 — will be echoed by many more companies as the year progresses. For workers whose roles encompass the repetitive layer of a professional function, the pressure to demonstrate AI fluency is becoming explicit on performance reviews and in job descriptions.

LinkedIn Fights the AI Spam

Notwithstanding the upward trend in corporate AI investment, there is a growing consumer and professional backlash against low-effort AI-generated content. In May 2026, LinkedIn announced it was expanding enforcement actions against AI-generated comments — specifically, comments posted across multiple posts using automation tools with minimal human involvement, including those that "restate the original post" without adding meaningful contribution. The crackdown followed a broader internal shift LinkedIn made earlier in 2026 to reduce the reach of content flagged as generic or repetitive. The platform's challenge reflects a wider issue: as AI-generated prose becomes indistinguishable from human prose in short-form content, social platforms are struggling to protect signal from endlessly amplified noise.

5. Security and Governance: Where AI Changes the Threat Model

The Take It Down Act Goes Live

A landmark AI governance development arrived in May 2026 when US enforcement authorities unsealed criminal complaints against two individuals accused of posting thousands of non-consensual deepfake intimate images online. The charges were brought under the Take It Down Act — legislation that had passed nearly a year earlier but whose criminal enforcement provisions only officially came into force on the day of the unsealing. The law requires platforms to take down non-consensual deepfake content within a specified timeframe or face civil penalties. The cases represent the first time criminal charges have been filed under this specific federal deepfake statute.

The arrival of actual enforcement is significant because it changes the legal risk landscape for anyone running visual generative AI products that can produce photorealistic human imagery. Before the Act's enforcement date, the legal exposure was essentially theoretical — platform liability frameworks had not yet caught up with technology. With the first prosecutions now public, downstream users and product owners face clear new obligations around user reporting workflows, proactive content scanning, and documented content takedown pipelines.

GitHub Breach, VS Code Extension Supply Chain

Another security story from May 2026 illustrates a broader structural vulnerability in the AI era's developer toolchain. GitHub disclosed that a data breach in May affected 3,800 of its internal repositories — traced to a "poisoned" VS Code extension installed on an employee device. The hacking group TeamPCP claimed responsibility. GitHub confirmed the exfiltration was limited to internal data, but the episode underscored a class of risk that has become dramatically more consequential as developer tooling becomes more dependent on local AI integrations, external model APIs, and plugin ecosystems.

The "poisoned extension" pattern — malicious code delivered through a developer-adjacent tool — is now listed alongside supply chain attacks on npm packages and Docker images as a serious, ongoing attack vector. Developers using AI-augmented code editors should consider running those tools in sandboxed environments, particularly on machines that hold access to production infrastructure or internal repositories.

Nvidia Driver Vulnerabilities

Nvidia also disclosed in May 2026 that drivers released before version 596.36 contained unspecified "high severity" security vulnerabilities affecting both Windows and Linux GPU driver installations, as well as its vGPU product. The update to 596.49 shipped on May 12, 2026 — after which point running the newest driver resolved the exposure. The episode was a reminder that the compute infrastructure layer — the GPUs, their drivers, the chip firmware — is a live software supply chain in its own right. As AI clusters grow larger and more exposed, the security baseline for that layer needs to be treated with operational rigour equal to that applied to the application stack.

6. The Gen Z AI Context Collapse

Young Adults Are Growing Tired, Resentful — and Sometimes Exiting the Industry

One of the most significant and under-covered technological stories in 2026 is the growing estrangement of an entire generation from AI. Gallup polling through late 2025 and early 2026 has tracked a dramatic shift in Gen Z attitudes toward AI tools: only 18% of 18–29 year-olds now say they are "hopeful" about AI — down from 27% in 2025 — and only 22% say they are "excited." A solid majority now say they believe AI's risks outweigh its benefits.

The Verge reported a longform profile in May 2026 that captured concrete cases: a 27-year-old art teacher in Los Angeles who has sworn off using any chatbot tool, not out of ignorance but because each use feels like it is hollowing out something irreplaceable in her professional practice. A 25-year-old cloud infrastructure engineer who left a major Silicon Valley AI infrastructure employer and left the tech industry altogether — becoming a food services worker in New York — citing environmental concerns about data centre energy use and ethical discomfort with the entire direction of the field. A consistent undercurrent through all these accounts: young people are being told on one hand that AI is essential futureproofing for their careers, and on the other that it will eliminate jobs, degrade human communication, and hollow out meaningful work. The cognitive dissonance is producing resentment, not abandonment.

Harvard and Gallup's study, a close look at which The Verge profiled, found that 74% of young adults in the United States use a chatbot at least once a month. But 79% expressed concern that AI makes people lazier, and 65% said chatbots "promote instant gratification, not real understanding." The contradiction is extreme: near-universal familiarity and regular use, paired with near-universal scepticism and growing resentment.

For companies and employers building strategies around AI, this demographic signal is important. In Australia, the UK, and many EU states, Gen Z workers are representing themselves to employers as technically skilled but professionally suspicious about AI integration. Some companies have started responding with AI-optional policies — announcements that permitted AI use will never count against performance assessments, and that humans will not be penalised for rejecting tool-assisted work in favour of demonstrating craft. This is a middle ground worth watching.

7. Cars and Transportation: The Quiet Electric Pivot

EV Adoption Slows — But Electrification Continues

While AI has dominated technology through 2025 and 2026, the consumer EV story has steadied into something quieter but structurally significant. The early 2020s saw explosive growth in consumer EV adoption globally, driven by government incentives, consumer enthusiasm, and aggressive auto brand electrification targets. Through 2026, adoption has slowed in mature markets — US EV sales growth softened for both Tesla and legacy brands — but not stalled. China's EV market, the world's largest by volume, continued its structural shift despite broader economic concerns. BYD's energy storage and vehicle platform verticalisation continued to give it a cost advantage over both legacy Western manufacturers and Tesla in mid-2026.

The more consequential trend in 2026 is the electrification of commercial and heavy-duty vehicles. Electric semi-trucks — Rivian's Amazon delivery fleet scaling, Tesla's Semi entering wider commercial production, and start-ups like Arrival — are moving from field trials to sustained commercial deployment. Several US states have adopted updated medium-duty EV purchase incentives as part of updated clean fleet rules. This undermines the common media narrative that EVs stalled. They stalled for consumer passenger cars in a few wealthy Western markets; for commercial and heavy-duty applications globally, the story is more like a quiet revolution is still underway.

Autonomous Driving: Waymo Grows, Tesla Restructures FSD

Autonomous driving — always the furthest-out bet in transportation tech — underwent a meaningful structural shift in 2026. Waymo expanded commercial service coverage to at least six additional US cities beyond Phoenix and San Francisco, and for the first time publicly disclosed profitability metrics for the ride-share business in its earliest operating markets. This is a significant moment: no autonomous vehicle company has even come close to a clean accounting of unit economics in ride-share, and Waymo's disclosure implies a path to operational sustainability anchored to a real commercial business model.

Tesla's Full Self-Driving (FSD) product platform meanwhile underwent a substantial internal restructuring in Q1 2026, with the company integrating end-to-end neural planning models directly into its vehicle firmware for all 2024+ models. The user-facing product remains supervised, with the company maintaining the explicit position that the human driver is responsible at all times. What shifted is the internal model architecture and the degree to which Tesla's self-driving stack is now downstream of the same neural network training ecosystem as the rest of its AI stack. We are approaching an era where the distinction between a "self-driving car" and a "car with AI co-pilot software" will stop being a category distinction and start being a regulatory one.

8. Biotech: CRISPR, mRNA, and the Protein-Folding Moment

CRISPR Goes From Engineering Tool To Medical Record

One of the most quietly consequential technological shifts in 2026 plays out almost exclusively in academic journals and biotech conference data rooms: the CRISPR gene-editing toolkit is being systematically upgraded from an experimental research tool into a clinically validated, commercially delivered set of medical therapies. By early 2026, at least two CRISPR-based gene editing therapies — one targeting sickle cell disease and the other targeting a beta-thalassemia form — had received full FDA approval and moved from clinical trial to standard insurance-covered care pathways. This is a landmark, not just in biotechnology, but in the history of medicine: for the first time, a disease-specific human genome correction therapy is a prescribed treatment, not an experimental procedure.

Beyond these early approvals, the CRISPR R&D pipeline in 2026 is extraordinarily broad. CRISPR mammalian-embryo editing — an ethically sensitive frontier — remains paused at clinical lines but is advancing in animal models for conditions including Huntington's disease and hereditary forms of familial hypercholesterolaemia. The foundational protein engineering work is now generating gene editors built on a family of "base editor" modifications that do not cut the DNA strand but chemically convert one base pair to another, an approach that eliminates one of the primary safety concerns around older CRISPR-Cas9 cleavage. A set of base-editor-based therapies targeting TTR amyloidosis moved into Phase 2 clinical trials in early 2026. The machine-learning-accelerated protein design breakthroughs first demonstrated by DeepMind in 2021 have now been productivity-multiplied by new AI-designed base-editor and prime-editor enzymes, massively accelerating the speed with which new gene editors can be developed for new targets.

The Protein-Folding Breakthroughs Quietly Advance

AlphaFold — DeepMind's AI system that predicted three-dimensional protein structures from amino acid sequences with accuracy that implied it had solved a 50-year-old structural biology problem — received a significant upgrade in 2026, with AlphaFold 3 substantially expanding predictive coverage beyond proteins to the full universe of biomolecular complexes (protein-ligand, protein-protein, protein-RNA, etc.). The Rosalind Franklin Institute and others began deploying AlphaFold 3 in real drug discovery pipelines: screening metabolic pathways, designing small molecules for targets that previously had no structure data, and accelerating hit-identification in oncology from the multi-month range to multi-week. The speed ration this imparts is comparable in scale to the efficiency gains in software that transformed coding in the 2010s, and is only now starting to be felt in biotech deal-making and drug development economics.

mRNA: Normalization and Platform Expansion

mRNA therapeutics — catapulted into mainstream awareness by the 2020 Pfizer-BioNTech and Moderna COVID-19 vaccines — have in 2026 normalised from a one-problem wonder into a continuing platform opportunity. Moderna in early 2026 reported Phase 3 trial results for its mRNA-based cancer vaccine personalised to each patient's neoantigen presentation. Early data was positive enough that the treatment began transitioning to community-cancer-centre deployment in late 2025, with first commercial doses administered in early 2026 in Australia and the northeastern United States. Pfizer also moved its own mRNA personalised cancer vaccine toward regulatory submission in Europe. The mRNA oncology platform — once speculative — is now in the early innings of real-world validation.

9. The Frame: What This All Means in Practice

The story threading through all these layers of technology is not a single narrative arc. It is a set of contradictions that together describe a field that has crossed an adoption threshold and is now navigating early post-hype dynamics:

• AI economics are hyper-profitable, but adoption friction is real. Nvidia and the hyperscalers are printing revenue at rates that infrared epochal change, yet the workplace applications of AI tools remain contested by the people who will have to use them — including younger workers who feel they were never given a choice.
• The open-source stack is catching up, and proprietary dominance is thinning. Chinese AI firms and open-weight models are now competitive with or better than top-tier proprietary products; this has not yet collapsed pricing, but it has begun to shift procurement conversations.
• Agents are the product category everyone predicted would exist eventually; 2026 is where "eventually" arrived. Figma's native agent, GitHub Copilot's wider codebase integration, and AWS/GCP/ Azure's agent orchestration platforms together describe a transition from AI as a tool to AI as a platform layer. That is a substantially different thing.
• The biotech revolution is real, but slower than the narrative suggests. Gene editing has crossed the threshold from experimental to reimbursable in one clinical indication, and mRNA has expanded beyond pandemic response to oncology. These are epochal for the patients involved and for healthcare economics globally — but they are not transforming the entire biotech or healthcare sector overnight.
• Security and governance risk scales with AI adoption. Every organisation investing in AI substantially also needs to be auditing the security of the underlying infrastructure, the supply chain of any extensions or plugins used alongside AI tooling, and its own compliance posture around AI-generated content and deepfake policy.

Conclusion

The framing around AI through 2025 and into 2026 was defined by two contradictory impulses: breathtaking investment and earnings on one side, and chronic cultural and workforce skepticism that grew faster than anyone projected. Both are real. Both matter. Neither invalidates the other.

The organisations and product teams building or adopting AI are navigating a complex set of pressures simultaneously: competitive urgency to move fast, regulatory caution to move in compliant and defensible directions, lasting scepticism from half the workforce, and a technology base that is genuinely improving even as the sociology around it is not catching up. That is a hard but solvable equation. 2026 is the year that equation stops being hypothetical and starts being solved in production.

For technology practitioners, engineers, designers, business leaders, and anyone following how work itself is changing — this is the most interesting moment since the original introduction of the web to commercial infrastructure. There remains enormous opportunity. It just comes with a complexity tax that the hype narrative did not prepare anyone for.

The companies that will succeed in the AI era are not necessarily the ones with the best models. They are the ones that navigate these contradictions most clearly: build technically sound products, treat their workers with genuine investment partnerships rather than pure cost-reduction framing, maintain a serious and documented approach to infrastructure and content governance, and build products on software platforms that do not collapse under load.

The year ahead is wide open. The era of AI-try-at-every-expense is ending. The era of AI-with-consequence is still being written — and that is exactly where the best work happens.