Insights // the AI wave, in numbers

Four years in, the numbers tell a calmer story than the headlines

Manta Labs ships AI work for working businesses. The brief on this page is the same brief we walk into every engagement with — what is actually moving, by how much, and where the money goes. We refresh it as new data lands.

As of 6 May 2026 · sources at the bottom

0,

Models tracked since 1950 (Epoch AI)

0.0T

Tokens processed every day across the major providers

0.00M

Hopper-class GPUs deployed worldwide by end of 2025

$0.0B

Private AI investment in the US, 2025 (≈23× China)

01

The labs

The revenue race

Two quarters ago this was a one-horse race. It isn't any more.

Annualized revenue//OpenAI · Anthropic · Google AI

Run-rate, not booked. Numbers come from disclosed quarterly figures and Epoch AI's revenue dataset. The yellow band marks the quarter where Anthropic crossed OpenAI on annualized run-rate.

$0.01B$0.10B$1B$10B$100BQ1 22Q1 23Q1 24Q1 25Q1 26OpenAI$24BAnthropic$30BGoogle AI$14BAnthropic passes OpenAI on annualized revenue.

Source: Epoch AI revenue tracker; SaaStr; Sacra.

$28M → $24B

OpenAI: roughly 850× in four years.

~$100M → $30B

Anthropic: a 300× climb, mostly in the last six quarters.

0 → $14B

Google AI: late to monetize, now compounding fast.

02

The training

The compute curve

Frontier models double on training compute every six months. They have done so for a decade. The slope hasn't bent.

102310241025102610272020202220242026GPT-3ChinchillaPaLMGPT-4Gemini UltraClaude 3 OpusLlama 3.1 405BGPT-5Grok 4Claude 4 Opus4–5× per year (Epoch AI)

Source: Epoch AI — large-scale AI models database. Y-axis is log-FLOPs.

Two practical implications. First, the largest run on record (Grok 4) emitted more than 72,000 tons of CO₂-equivalent — published in the 2026 AI Index. Second, the next jump on this chart will not come from more H100s; it will come from the GB200 / GB300 generation now landing in hyperscaler racks.

03

The price

The cost of intelligence collapsed

A thousandfold cheaper, in three years, for the same work. This is the chart that matters for anyone shipping product.

1,000×

Cheaper to run GPT-3-class intelligence today than at launch. GPT-4-class follows the same curve, eighteen months behind. Token prices fall roughly an order of magnitude each year.

Nov 2022 → May 2026 · per million tokens

$0.01$0.1$1$10$100Nov 22May 26GPT-3.5-class$0.008/MGPT-4-class$0.400/M

Source: Epoch AI — LLM inference price trends; Stanford AI Index 2026.

04

The volume

The token tsunami

If revenue is the score, tokens are the play count. One provider is now larger than the next four combined.

0T50T100T150T200T250T20232024 H12024 H22025 H12025 H22026 H1ByteDance180T/dayGoogle22T/dayMicrosoft18T/dayOpenAI12T/dayAnthropic9T/day

Source: provider disclosures (Microsoft, Google), OpenRouter analytics, ByteDance press notes.

ByteDance's lead is almost entirely AI-generated short video — a different use case than the chat-and-code workloads driving Western providers. For the western stack, agentic flows are the new headline; one well-scoped agent run consumes more tokens in an hour than a whole team of users did in 2024.

05

The bill

The electricity bill is real

The IEA expects data-centre demand to roughly double by 2030. AI is the reason.

0 TWh250 TWh500 TWh750 TWh1000 TWh20202022202420252026E2030E320460415485580945AI workloadsOther data-centre load

Source: IEA — Energy and AI (2025 + 2026 updates). Yellow segments are the AI share of total data-centre load.

+50%

AI-focused data-centre electricity, 2025 alone.

945 TWh

Projected total data-centre electricity in 2030 — about 3% of global demand.

Expected growth in AI-optimised data-centre electricity by 2030.

06

The market

Adoption faster than the internet

Generative AI hit 53% population adoption in three years. The internet took seven. The PC took ten.

0%15%30%45%60%launch+1y+2y+3y+4y+5y+6y+7y+8y+9y+10yGenerative AI53% at +3yThe internet51% at +10yThe PC31% at +10y

Source: Stanford AI Index 2026 (consumer adoption); Pew Research; historical Census Bureau.

07

The releases

Four years of frontier

Yellow dots mark releases that reset the frontier. Everything else is the field keeping up.

2023202420252026ChatGPTOpenAIGPT-4OpenAILlama 2MetaClaude 2AnthropicClaude 3.5 SonnetAnthropicGemini 1.5GoogleLlama 3.1 405BMetaDeepSeek R1DeepSeekGrok 4xAIGPT-5OpenAIClaude 4 OpusAnthropic

Source: Epoch AI; llm-timeline.com; provider announcements.

How this stays current//the boring middle

We refresh this page from a pipeline, not a press cycle. An Inngest job runs once a day, pulls the latest CSVs and APIs from the sources below, normalizes the fields the page renders, and writes a snapshot to Postgres. The page reads the latest snapshot at request time. If a source is down or a number looks wrong, the previous snapshot stays live.

See site/app/(marketing)/insights/DATA_PIPELINE.md for the implementation notes.

If the chart that matters most is the cost-collapse one, the question is what you ship with it

That's where we come in. Pick a workflow that's bleeding time. We'll ship a working fix.

Insights — the AI wave, in numbers · Manta Labs