Hyperscalers will spend $710B on AI capex this year. The cert system is broken. The junior pipeline is breaking. Two thirds of enterprise AI teams already run MCP-backed agents in production. The transition is not coming. It is happening now. Here is the map.
Year-by-year: what happens, what it changes, and what to watch for to know if the prediction is breaking.
The Coding Agent Becomes the Default IDE. Per JetBrains' April 2026 developer survey, Claude Code at 18% work adoption (1.5x the September 2025 figure, 6x the April 2025 figure) and Cursor at 24% primary-tool selection have moved coding agents from "interesting tools" to standard issue. GitHub Copilot still leads on awareness (76%) but its primary-tool share is contested. Most pro developers run a three-tool stack rather than committing to one.
MCP becomes infrastructure. The Model Context Protocol — Anthropic-originated, openly governed — passes 78% adoption among enterprise AI teams running at least one MCP-backed agent in production (April 2026 survey). Anthropic, OpenAI, Google, Microsoft, Salesforce, Cloudflare, and Replit have all shipped first-class MCP support. The "USB-C for AI tool use" framing is now reality.
Capex passes the trillion-dollar collective threshold. Big Four hyperscaler AI capex hits $710B for 2026 (Amazon ~$200B, Microsoft ~$120B, Google ~$185B, Meta ~$135B). Goldman Sachs models the 2025-2027 envelope at $1.15T. Capex as percent of revenue is now in the 45-57% range, territory that historically belongs to telcos and utilities, not software companies.
The agent layer arrives at the application boundary. What 2025 called "agentic workflows in pilot" become standard production pieces of incident response, code review, customer support triage, and data pipeline operations. Not because of any single launch. Because the substrate (MCP, frontier models with reliable tool use, open-source orchestration) has reached "boring" status.
The middle of the IT org chart starts hollowing. Junior developer hiring slows. L1 helpdesk headcount per employee ratio drops measurably (the data is messy because companies redirect-rather-than-fire). Platform engineering teams expand to absorb the coordination work. The prediction here is not "everyone gets fired." It is "fewer people do more, with the missing layer redirected to higher-leverage work or pushed out of IT entirely."
Counter-signal to watch for: if frontier model improvement slows in late 2026 (smaller capability deltas between Opus 4.7, GPT-5.5, Gemini 3.1), the agent-as-default trajectory delays by 12-18 months. The market priced in exponential capability gains. Linear gains break a lot of plans.
The certification system snaps. Frontier models pass every multiple-choice cert exam at human expert level. Cert bodies pivot to performance-based assessment under timed lab conditions or lose hiring signal entirely. Cisco's hands-on CCIE tradition becomes the model everyone copies. AWS, CompTIA, ISC2, and Microsoft are all piloting performance-based assessment by 2027. By 2028 it is mandatory or the cert is dead.
EU AI Act enforcement creates the first compliance moat. The Act is fully applicable as of August 2, 2026; penalties for general-purpose AI providers reach 3% of global turnover or €15M, whichever is higher; the Commission has exclusive Chapter V enforcement authority. By 2028, compliance posture is a real competitive variable. Anthropic's safety-first positioning shifts from marketing line to regulatory arbitrage. Open-source providers face a structural choice: meet the systemic-risk thresholds or stay below them.
VMware exodus completes. Broadcom's 300-1200% price increases on VMware (AT&T's bill went from $30M to over $100M; Gartner reports Proxmox VE evaluations up 340% YoY) drive a multi-year migration that lands hardest in 2027-2028. Proxmox, Nutanix AHV, and OpenShift Virtualization absorb most of it. Some enterprises repatriate to bare-metal Kubernetes. The "VMware admin" specialty becomes a niche legacy skill.
The two-bloc compute reality calcifies. US export controls have been in force since 2022 and tightened repeatedly. China's domestic stack — Huawei Ascend, SMIC fabrication, DeepSeek/Alibaba/Baidu/Moonshot models trained domestically — is no longer a curiosity. DeepSeek V4 shipping in April 2026 optimized for Ascend rather than Nvidia (reportedly at Beijing's direction) was the inflection. Multinationals navigating both blocs face genuine architectural divergence in their AI stacks, not just vendor choice.
FinOps becomes a P&L-defining function. AI workloads have unbounded cost surfaces in a way classical compute did not. A single prompt can scale from $0.0001 to $10 depending on context, model, and reasoning depth. By 2029, an enterprise without a senior FinOps leader who deeply understands inference economics is paying 2-3x what a comparable competitor pays. This is no longer a Cloud Architect's side project.
Post-quantum cryptography migration becomes the largest IT project in regulated industries. NIST's PQC standards (FIPS 203, 204, 205, finalized August 2024) drive mandatory migration timelines for federal contractors and most financial services. "Harvest now, decrypt later" attacks are confirmed in the wild — nation-states have been collecting encrypted traffic for years. The migration is an 8-figure project for any large enterprise. Engineers who understand both classical and PQC stacks are the rarest hire in security.
AI-native operations is the default architecture for new systems. "Cloud first" (the 2012 doctrine) is replaced by "AI-native first": every new system is designed assuming AI agents operate, monitor, and extend it. APIs everywhere, structured logs everywhere, observability as a first-class concern. The human exception path is designed last, not first. The stragglers are not new builds — they are the enterprise legacy moving from cloud-migrated to AI-operable, which is genuinely hard work.
The IT generalist is extinct above 200 employees. Specialization is total. The 2030 enterprise IT org chart has roughly four functions: Platform Engineering (the substrate), AI Infrastructure (the model and agent layer), Security & Governance (the controls), and Data/ML Engineering (the pipeline). Traditional Operations is a small fleet-management team running AI agents. Helpdesk that survives is high-empathy human escalation; everything else is automated.
Programming language consolidation is real but not absolute. Python (AI/ML/data, locked in), TypeScript (web, locked in), Rust (systems, displacing C/C++ in security-sensitive domains), Go (backend services), and SQL hold. Java survives in enterprise legacy with declining new adoption. Everything else is niche. A new AI-native language (early candidates: Mojo from Modular) may emerge but adoption takes years.
The new normal. IT is no longer a department, it is the substrate. Every company is a tech company, every tech company runs on AI infrastructure, and the question "do we have AI in this product" is as quaint as "do we have a website" was in 2010. The total IT workforce is smaller; per-capita compensation in surviving roles is meaningfully higher; the bottom quartile of 2025 IT roles has moved into adjacent fields (data labeling, compliance, technical sales) or out of tech entirely.
Infrastructure is ambient. AI is ambient. Frontier model capability has plateaued or grown logarithmically in the late 2020s; the action has moved to distribution, integration, and economics. The companies that win are not the ones with the smartest model but the ones with the deepest integration into how work happens. Microsoft (via Office and GitHub), Google (via Workspace and Android), and Amazon (via the cloud floor) are the structural winners regardless of who has the best model in any given quarter.
What we will be wrong about: if cryptographically relevant quantum arrives sooner than 2030 (low probability but rising), the entire timeline accelerates and shatters, especially in security. If frontier model gains stop entirely after 2027 (unclear but possible), the agent layer arrives but does not deepen, and many of the headcount predictions soften. Calibration over confidence: these are bets, not prophecies.
Eight sectors, what the work looks like by 2030, what dies, what is born. Bullets are predictions with horizon 3-5 years.
Confidence reflects signal stability, not enthusiasm. "Very high" means the structural advantage is durable for several years. "Med" means real uncertainty; the prediction could flip on a single major event.
| Race | Leader 2027 | Leader 2029 | Leader 2031 | Confidence |
|---|---|---|---|---|
| Frontier Models | Anthropic / OpenAI (alternating quarters) | Three-way tie or Google pulls ahead | Plateau; Google long game pays off | MED |
| Inference Hardware | NVIDIA (>85% share) | NVIDIA + hyperscaler custom silicon | Fragmented; NVIDIA still largest but not monopoly | HIGH |
| Enterprise AI Platform | Microsoft (Azure + Copilot) | Microsoft + Google split | Distribution-based; bundled wins, standalone loses | HIGH |
| Open Source Models | DeepSeek, Qwen, Llama 4 | Chinese labs lead by capability/cost | Permanent commodity layer | VERY HIGH |
| Coding Agents | Anthropic (Claude Code) + Cursor | Stack consolidation around 2-3 winners | Agents are bundled into IDEs and OS | HIGH |
| Sovereign / National AI | US, EU, China each scramble | Three-bloc reality: US, China, EU+allies | Politically permanent | VERY HIGH |
| MCP / Tool Use Standard | MCP wins (78% enterprise adoption) | MCP + competing standards consolidate | Universal standard, OpenAPI for the LLM era | HIGH |
Verdict, then reasoning. Power meters indicate present-day strength on AI capability and enterprise distribution. Verdicts are 5-year horizon, not next-quarter calls.
Does not need to win the model race. Owns Office (400M+ paid seats), GitHub (the developer substrate), Azure (#2 cloud with deep enterprise trust), Windows, and Teams. Copilot is shipped into every surface. The OpenAI partnership gives them frontier access without exclusive dependence; in-house models (MAI series) provide insurance. Risk: Google catches up in productivity AI and large customers split workloads. Hardest moat to displace in big tech.
Best raw cards: TPUs (vertically integrated silicon), Search and YouTube data, DeepMind talent and culture, Gemini reaching frontier parity, Workspace and Android distribution. Gemini 3 Pro tops LMArena at 1501 Elo (early 2026). Historic weakness is execution, not capability — the brand-caution and product-fragmentation tax has been real. The 2028-2031 window is Google's to win or fumble.
Default cloud for the bulk of Fortune 1000. Bedrock provides model-agnostic access without betting on one provider. Trainium and Inferentia silicon are credible Nvidia alternatives at hyperscale. Will not win frontier AI but does not need to — they win by being the floor everything runs on. Anthropic investment ($8B+ committed) is structural. Amazon plus Anthropic is one of the strongest non-Microsoft AI plays in enterprise.
Punches well above weight. Claude Opus 4.7 (April 2026) leads on agentic coding and computer use; Claude Code at 18% developer-work adoption with 46% "most-loved" satisfaction in JetBrains' April 2026 survey. MCP origination gives them protocol leverage. Safety-first positioning is regulatory arbitrage if EU AI Act enforcement bites. Risk: small organization, capital-constrained vs hyperscaler partners, dependent on AWS for infrastructure. The bet pays asymmetrically.
$193.7B annualized datacenter revenue, ~90% of AI accelerator spend. CUDA is still the deepest moat in tech but is being tunneled from every direction: AMD MI300/MI400 series, Google TPU v6, AWS Trainium 2/3, custom silicon at every hyperscaler, Cerebras and Groq for specific workloads. PyTorch 2.x is increasingly hardware-agnostic. NVIDIA's 2031 position is real but reduced from today's near-monopoly. The capital-allocation thesis is "high but compressing margins."
Llama 4 (April 2025) was the open-source frontier; Llama 4 Maverick is competitive with GPT-4o-era closed models. Then Meta shipped Meta Muse Spark closed-weight in April 2026 — a real departure from the three-year open strategy. Mark's bet appears to be commoditize-when-behind, capture-when-ahead. AI capex of $135B in 2026. They do not need enterprise wins; they need social and ad-targeting dominance, where AI is the engine. Competitive with Llama 5+ open releases TBD.
Apple Intelligence (announced 2024) is the most privacy-preserving consumer AI deployment at scale. M-series silicon is exceptional for on-device inference. Strategy is ambient, private, on-device AI that does not require sending personal data anywhere. If privacy regulation tightens, Apple's architecture becomes a structural advantage. Siri remains a chronic execution embarrassment. The hole in the plan is real and persistent. Apple's AI story is "patient capital" not "decisive lead."
DeepSeek V4 (April 2026) ships at $0.14/$0.28 per million input/output tokens for V4-Flash with reasoning capabilities comparable to mid-2025 frontier models, optimized for Huawei Ascend silicon. Alibaba's Qwen, Moonshot's Kimi, Zhipu's GLM round out the Chinese frontier. Cannot sell into US federal or most large US enterprise, but exports massive cost pressure that all other vendors must match. Permanent feature of the landscape.
WatsonX is technically credible for regulated industries. The brand keeps fumbling the narrative. Revenue ceiling is low and the margin profile is services, not software. IBM survives as a government and financial services incumbent with Red Hat as the genuinely valuable software asset. Does not shape the future of IT; participates in it.
Real surprise upside on AI infrastructure: OCI is a credible NVIDIA-dense cloud, used heavily for foundation model training (OpenAI's $300B Stargate-related contract is a real signal). But Oracle's brand is "expensive and adversarial" and CIOs know it. Database business is perpetual. Cloud business will not catch AWS or Azure. The Ellison AI data center pivot is real infrastructure; whether Oracle captures the upside or just becomes a GPU landlord is open.
These are the questions that come up in private and get hedged in public. Answers are calibrated bets, not certainties.