2025 was the year renewables overtook coal in global generation — Ember's April 2026 Global Electricity Review logged 33.8% renewable share vs 33.0% coal, the first time in 100 years coal was displaced. Total installed renewable capacity reached 5,149 GW (IRENA, April 2026), up from 3,870 GW end-2024. Net adds hit a record 692 GW — solar PV alone accounted for ~505 GW of that. Energy-transition capital came in at USD 2.3 trillion (BNEF), up 8% year-on-year, with climate-tech VC/equity surging 53% to USD 77.3 bn. Renewables + wind + hydro avoided ~2.6 Gt CO₂e in 2025 versus a fossil counterfactual.
Solar PV is now the dominant growth engine: ~505 GW added in 2025 alone, pushing cumulative capacity past 2,370 GW. Onshore wind adds held at ~140 GW/year. Offshore wind underperformed its 2024 target (8 GW added vs forecast ~12 GW) on rate pressure + permitting slowdowns, but the 2025 pipeline under construction rebuilds momentum (33 GW in build globally — China 9.1 GW, UK 7.6 GW, US 5.9 GW). Battery + thermal storage are the compounding frontier: battery doubled again year-on-year.
| Technology | 2020 | 2022 | 2024 | 2025 | Note |
|---|---|---|---|---|---|
| Solar PV | 714 | 1,055 | 1,865 | ~2,370 | China past 1,000 GW first time |
| Onshore Wind | 697 | 838 | 1,027 | ~1,170 | Stable pipeline; normalised supply |
| Offshore Wind | 34 | 64 | 80 | 89 | 33 GW under construction |
| Hydro | 1,211 | 1,256 | 1,283 | 1,301 | Mostly pumped-storage modernisation |
| Bioenergy | 128 | 149 | 156 | ~160 | Slow; supply-chain bottlenecks |
| Battery storage (GWh) | 17 | 58 | 260 | ~500 | 92 GW / 247 GWh added in 2025 |
| Thermal storage (GWh-th) | ~234 | ~260 | ~330 | ~380 | Preliminary; Rondo + Antora scaling |
Source: IRENA Renewable Capacity Statistics 2026 for solar / wind / hydro / bio; BNEF Global Energy Storage 2025 for battery cumulative; IRENA Innovation Outlook (2020) extrapolated for TES — no global TES register yet exists.
Total energy-transition investment hit $2.3 trillion in 2025 — up 8% year-on-year. The mix has shifted materially: electrified transport ($893 bn) is now the single largest category, pushing past renewable generation ($690 bn) as the transition moves from building new supply to electrifying end-use. Grid investment ($483 bn) is finally catching up with generation — a long-flagged bottleneck. Climate-tech VC/equity surged to $77.3 bn (+53% YoY) and transition debt issuance hit $1.2 tn (+17%).
China is now ~42% of global renewable capacity (2,159 GW), underpinned by NDRC Document 136 (market-pricing shift) and the first country ever past 1,000 GW of solar mid-2025. EU at ~18% (934 GW) tracks 66% RE-electricity share — short of the 69% REPowerEU target but moving. US at ~9% faces headwinds from Trump EO 14154 (Jan 2025 IRA freeze, 781 EPA grants cancelled) though legal-impoundment-doctrine challenges limit full rollback. India ~5% (254 GW total RE, 133 GW solar) — hit its 50% non-fossil capacity target five years early; PM Surya Ghar rooftop scheme delivered 7 GW across 2.4M households by end-2025.
| Region | Share | Capacity | Key driver |
|---|---|---|---|
| China | ~42% | 2,159 GW | NDRC Doc 136 market-pricing shift · first past 1 TW solar |
| European Union | ~18% | 934 GW | REPowerEU · Clean Industrial Deal €100bn · EU ETS signal |
| United States | ~9% | ~460 GW | IRA operative but under freeze (EO 14154, Jan 2025) |
| India | ~5% | 254 GW (133 solar) | 500 GW non-fossil by 2030 · PM Surya Ghar · CCTS mid-2026 |
| Rest of Asia | ~12% | ~620 GW | Japan GX · Korea K-ETS · Vietnam PDP8 grid |
| Rest of world | ~14% | ~720 GW | Brazil hydro + COP30 solar push · MENA auctions · Africa <3% |
Once renewable penetration crosses ~55%, storage + grid flexibility become the binding constraint — which is why dispatchable heat (TES) and grid batteries are re-pricing on dispatchability, not just tCO₂e. Battery storage installed 92 GW / 247 GWh in-year in 2025, taking cumulative past ~500 GWh. BNEF expects ~15× growth to ~2 TW / 7.3 TWh by mid-2030s. Thermal energy storage is earlier but scaling: Rondo (brick, 1,000–1,500°C, 16–18hr discharge) now operates commercial units at Calgren CA and is expanding into Indian cement + steel; Antora (graphite, up to 2,000°C, 25hr) commissioning Fresno CA facility 2025.
Main-case trajectory: 9,530 GW cumulative by 2030 (2.6× 2022), generation share ~46%. Accelerated case: 10,400+ GW (~90% of the COP28 tripling pledge, ~50% generation share). COP28 tripling target of 11,500 GW remains off-track in main case. Post-2024 policy shifts: Trump IRA freeze modelled as slower — not reversed — US buildout; India National Electricity Plan 2023 (NEP13) draft pushes coal phase-down; EU REPowerEU three-years-on review (2025) maintained 592 GW solar + 510 GW wind by 2030 targets.
Industrial process heat consumes over 70% of industrial energy. AI-optimised CSP can abate ~10% of global CO₂ from IPH.
CAGR 15.2% (Precedence Research)
HVAC ≈ 40% of building energy. Generative "virtual engineer" agents unlock 25–40% savings.
CAGR 25.7% (MarketsandMarkets)
Early ReFi pooled legacy credits into fungible tokens (e.g., BCT), masking quality. Registries (e.g., Verra) reacted by halting unsanctioned tokenisation — credits now flow through curated, non-fungible architectures.
Project-specific assets preserve quality + enable per-project price discovery. No pooling blur.
Ex-ante financing funds new climate action — not legacy credits dredged from earlier vintages.
AI + DePIN streams automate certification — full auditable trails from sensor to issuance.
| Feature / Service | TenzaOne | Sylvera / Pachama | Regen Network | Toucan / KlimaDAO |
|---|---|---|---|---|
| Core business model | End-to-end financing + verification for new high-integrity projects. | Ratings SaaS for existing credits. | Protocol / registry for eco-credits. | Liquidity / marketplace for bridged credits. |
| AI-powered verification (dMRV) | ✔ | ✔ | ✖ | ✖ |
| DePIN / IoT integration | ✔ | ✖ | ✖ | ✖ |
| Finances future credits (ex-ante) | ✔ | ✖ | ✔ | ✖ |
| India CCTS path | ✔ ACV-Agency-compatible dMRV | ✖ | ✖ | ✖ |
| Asset type | Non-fungible (project-specific) | N/A (ratings) | Fungible + non-fungible | Fungible (pooled) |
| Investor tools (LLM, dashboards) | ✔ | ✖ | ✖ | ✖ |
Generative AI has moved from demo to production at a pace that shocked most analysts. 92% of Fortune 500 companies are now using OpenAI products (McKinsey Nov 2025) and 88% of organisations use AI in ≥1 function. VC investment hit a $87 billion record in 2025 for GenAI alone (EY / PitchBook), with total AI VC reaching $202 B — roughly 50% of all global VC. The revenue flip nobody expected: Anthropic surpassed OpenAI in early 2026 at ~$30 B ARR vs OpenAI's ~$24 B. Agentic AI is the next inflection.
Top models now score above ~90% on MMLU, which was the gold-standard 2023–2024 benchmark. Attention has shifted to harder tests. Per Artificial Analysis Intelligence Index (April 18, 2026), there is a three-way tie at 57 between Claude Opus 4.7, Gemini 3.1 Pro Preview, and GPT-5.4 (xhigh) — each strongest in different categories.
| Benchmark | Focus | Leader | Score | Runners-up |
|---|---|---|---|---|
| GPQA Diamond | Graduate-level science reasoning | Gemini 3.1 Pro Preview | 94.1% | GPT-5.4 92.0 · GPT-5.3 Codex 91.5 · Opus 4.6 91.3 |
| Humanity's Last Exam | Hardest multi-domain | Gemini 3.1 Pro Preview | 44.7% | GPT-5.4 41.6 · GPT-5.3 Codex 39.9 |
| SWE-Bench Verified | Coding agents | Claude Opus 4.7 | 87.6% | GPT-5.3 Codex 85.0 · Opus 4.5 80.9 · Gemini 3.1 Pro 80.6 |
| SWE-Bench Pro | Long-horizon coding | Claude Opus 4.7 | 64.3% | #1 by wide margin |
| GDPval-AA (agents) | Multi-step agent tasks | Claude Opus 4.7 | 1,753 ELO | Gemini 3.1 Pro 1,314 ELO |
| SciCode | Scientific code reasoning | Gemini 3.1 Pro | 58.9% | — |
| AIME 2025 | Math Olympiad | Grok 3 / frontier models | 86.7%+ | Now >90% for most frontier |
| MMLU (saturated) | Legacy benchmark | All frontier | ~92% | Industry moved on |
Pattern: no single winner — Anthropic leads agentic / coding; Google leads scientific reasoning + multimodal; OpenAI leads breadth / tool-use. Scores indicative; Artificial Analysis, LMSys Chatbot Arena, and Scale AI SEAL leaderboards are the live-tracking sources.
| Model | Release | Context window | Key capability jump |
|---|---|---|---|
| Llama 4 Scout/Maverick/Behemoth | 5 Apr 2025 | 10 M / 512 K / — | First native multimodal MoE open-weight (17 B active / 109 B – 2 T total) |
| Grok 3 | 17 Feb 2025 | 1 M | Reasoning + DeepSearch; trained on 200 K H100s |
| GPT-5 | 7 Aug 2025 | 272 K in / 128 K out | Unified reasoning + general model; "thinking" modes |
| Gemini 3 Pro | 18 Nov 2025 (preview) | Long-context SOTA | Best multimodal; agentic "vibe coding"; 81% MMMU-Pro |
| Claude Opus 4.5 | 24 Nov 2025 | — | Coding + spreadsheet leap |
| GPT-5.2 | Dec 2025 | 400 K (256 K effective) | Long-context + agentic tool-calling |
| Claude Opus 4.6 | 5 Feb 2026 | — | "Agent teams"; Claude in PowerPoint |
| Gemini 3.1 Pro | 19 Feb 2026 | — | Multimodal + GPQA / HLE leader |
| Claude Opus 4.7 | 16 Apr 2026 | — | Leads SWE-Bench; Claude Managed Agents beta same week |
| OpenAI Codex Background Computer Use | 16 Apr 2026 | — | Parallel macOS desktop agents |
Status flags: GPT-5 confirmed released. "Claude Mythos Preview" circulating on third-party leaderboards is not an Anthropic-announced public model as of April 21, 2026 — treat as rumour until confirmed. Post-4.7 Claude versions unverified.
| Deloitte 2026 stage | % |
|---|---|
| Transforming — new products / reinventing core processes | 34% |
| Redesigning key processes around AI | 30% |
| Surface-level use — little workflow change | 37% |
Shift vs 2024: deep-transformation cohort tripled (<10% → 34%);
"no plans" shrank from ~12% → rounding error.
McKinsey split: 88% adoption but only 6–7% "actually winning"
(high-performer bucket with enterprise-level EBIT impact); only 39% report EBIT uplift at
the enterprise level.
Agentic systems are moving beyond chat into autonomous task execution. The governance gap is the defining 2026 constraint: Deloitte 2026 finds only 21% of companies deploying agents have a mature governance model. No function exceeds ~10% "fully scaled" agentic deployment. Top use cases: software engineering (24% in tech), IT ops (22%), research / knowledge work, customer service, product development (18%).
The closed / open gap collapsed practically in 2025. Six labs now ship competitive open-weight models: Meta (Llama 4), Alibaba (Qwen 3.6), Google (Gemma 4), Mistral (Small 4 / Large 3), Zhipu (GLM-5.1), DeepSeek (V4). Closed-source still dominates US / Fortune 500 production spend (~70–80%); open-weight dominates self-hosted / regulated / APAC (China, India, Eastern Europe) + edge use cases.
"Identity / audit / approval loops" become a product category. SOC-2-for-agents certifications; first enforcement under EU AI Act + US state laws.
Text + image + audio + video native. Video understanding (Gemini 2.5+, Sora-class video-in) enterprise-grade. 81% MMMU-Pro (Gemini 3 Pro).
Apple Intelligence (Gemini licensing deal Jan 2026 via Private Cloud Compute), Gemini Nano on Pixel / Galaxy / Chromebooks, Phi-4 on Copilot+ PCs. Real privacy-preserving agentic AI on-phone.
DeepSeek V4 + Qwen 3.6 drive further inference-cost collapse; GPT-4-class performance becomes effectively free at the margin. Inference no longer the bottleneck.
Domain-specific: Med-Gemini (healthcare), Harvey (legal, most AmLaw 100), BloombergGPT-2 / OpenAI Hiro Finance, ClimateBERT (climate risk disclosure).
Cursor (Anysphere) valuation >$10 B; Devin (Cognition); GitHub Copilot multi-model routing; JetBrains Mellum. Coding tasks the biggest frontier-model spend line.
Enterprises no longer single-vendor. Portkey, OpenRouter, Vercel AI Gateway = table stakes. Cost / capability routing per workload.
Gap in the map. No frontier-scale climate-native LLM yet — whitespace opportunity for vertical players with domain data + dMRV experience. Direct TenzaOne angle.
The single clearest gap in the 2026 vertical-LLM map is climate / energy. Healthcare has Med-Gemini, legal has Harvey, finance has BloombergGPT-2 and OpenAI's Hiro Finance acquisition — but there's no frontier-scale climate-native LLM. Climate-specific stacks remain bolt-ons (ClimateBERT for risk disclosure, not an agent) rather than native intelligence. TenzaOne's AI assessment stack sits squarely in this whitespace.