I weighed the home pages of 39 AI companies: median 3.1 MB, and one 7.4 MB PNG on Replicate's front page

AI companies sell speed: faster models, faster inference, faster developers. I wanted to know how fast their own front doors are. On 4 November 2026 I measured the home pages of 39 AI companies: model labs, inference clouds, vector databases, AI coding tools and automation platforms.

Disclosure: I built the tool used here, the Page Weight & Carbon Checker on Apify. Method and how to rerun it are at the end.

How it was measured

For each home page the checker downloads the HTML and every resource the HTML references: scripts, stylesheets, images (one candidate per srcset), fonts from @font-face, icons, preloads and media. It records the bytes actually transferred (after gzip/brotli), counts requests and third-party hosts, and estimates CO2 per view with the Sustainable Web Design model (v4, 494 g/kWh grid intensity).

Two caveats, both important:

• No JavaScript is executed. Anything a script loads later (chat widgets, analytics beacons, client-rendered images) is not counted. For script-heavy sites the real weight is higher.
• Everything referenced is counted, including images a browser might lazy-load only when you scroll. So for image-heavy pages the number is closer to "full scroll" than to "first screen".

Treat the results as a consistent yardstick across sites, not as a Lighthouse score.

The numbers

• Median: 3.14 MB per home page, mean 4.63 MB
• Median requests: 99
• Median estimated CO2: 0.47 g per view
• Only 9 of 39 stay within a 2 MB / 60-request budget

Where the bytes go, summed over all 39 pages:

JavaScript gets the blame in most performance talks. On these pages it is the third-largest item, behind images and fonts.

The heaviest five

The pattern: the heaviest pages are not heavy because of frameworks. They are heavy because of one or two files that nobody resized. Replicate's 7.42 MB image and Stability's 17.6 MB image are PNGs; as WebP or AVIF at display size they would typically be a few hundred kilobytes.

DeepL is the interesting case. Its HTML, CSS and JavaScript are lean (0.84 MB of JS), but a full traditional-Chinese font is referenced from the English home page. Font subsetting or unicode-range would mean the file is only fetched when Chinese text is actually shown.

The lightest five

Groq is the nice one to point at: an inference company whose home page is under a megabyte.

The model labs sit in the middle: anthropic.com 2.07 MB, x.ai 1.28 MB, mistral.ai 2.99 MB, huggingface.co 3.15 MB. openai.com is not in the sample because it did not serve its home page to an automated client.

Third parties

The most common third-party hosts across the 39 pages:

• Google Tag Manager: 16 sites
• jsDelivr CDN: 9
• OneTrust cookie consent (cdn.cookielaw.org): 6
• Sanity CDN, Webflow assets, Cloudflare Insights: 4–5 each

Third parties add requests more than bytes here: the heaviest pages are heavy because of their own images, fonts and video.

Four fixes that cost nothing

1. Convert hero images to WebP/AVIF at display size. 30 of 39 pages here serve at least one PNG, JPEG or GIF; on the two heaviest pages images are over 90% of all bytes.
2. Subset fonts. Ship only the glyphs you use, and split large CJK fonts with unicode-range so they are fetched only when needed.
3. Don't autoplay a 10 MB video above the fold. Use a poster image and load the video on interaction, or encode a much shorter loop.
4. Set a page budget in CI. A 2 MB / 60-request budget would have caught every one of the top five before it shipped.

Carbon, briefly

With the Sustainable Web Design model, the median home page here causes about 0.47 g of CO2 per view; the heaviest is around 4.1 g. Per view that is tiny. A home page with a million views a month at 4 g per view is about 4 tonnes of CO2 a month, though, and on the heaviest pages most of it comes from images that could be far smaller. The model is an estimate with wide error bars, but the ranking between sites is robust because it is driven by bytes.

How to rerun this

curl -X POST "https://api.apify.com/v2/acts/fast_data_asap~page-weight-carbon-checker/run-sync-get-dataset-items?token=YOUR_APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://groq.com/", "https://replicate.com/"], "budgetTotalKb": 2048, "budgetRequests": 60}'

Each row has bytes by type, request counts, third-party domains, the five largest resources, uncompressed text files, images not in modern formats, poorly cached resources, CO2 per view and a budget pass/fail. You can also pass a sitemapUrl to check a whole site. Pricing is $0.004 per page; the 39 pages here took under two minutes. Failed pages are free.

Data: home pages of 39 AI companies, fetched once from a cloud server without running JavaScript, 4 November 2026. Sites change often; numbers are a snapshot.