Platforms

Exa says its agents wrote kernels that cut indexing cost 47% on Amazon's chips, not Nvidia's

Exa's first public number on the cost of running its index was earned on Amazon's AI chips rather than Nvidia's, and it lands two weeks after DeepSeek, the Chinese AI lab, set a 75-cent floor under the price of search.

Richard TangRichard Tangthree-minute read
Share
Will Bryk, co-founder and chief executive of Exa, smiling in a light blue button-down shirt against a plain wall
Will Bryk, Exa's co-founder and chief executive. Exa published its first public cost number for its web index on November 18, run on Amazon's Inferentia2 and Trainium2 chips.

Exa has published the first number on what its search index costs to run, and AI agents wrote the custom kernels behind it. On November 18, Exa, whose search API is what an AI agent calls for the live web, posted the details on AWS's machine-learning blog 1. It runs the embedding step of its web index on Amazon's−4.01% — Amazon, down 4.01 percent today Inferentia2 and Trainium2 chips rather than Nvidia−3.64% — Nvidia, down 3.64 percent today GPUs, and reports, in its own words, "a 47% improvement in performance and cost per 1 million documents embedded" 1. The figure is self-benchmarked, on a page co-published by Exa's cloud vendor 1.

A model change forced the rewrite

A better document model required re-embedding 30 billion high-priority documents, which temporarily doubled the embedding load, and Exa used that window to re-pick its silicon 1. Off-the-shelf kernels from Amazon's Neuron SDK came first, then two custom pieces 1. One fused kernel stopped the chip's memory stalls; the other rebuilt attention around the exact shapes of Exa's model, so no compute went to padding 1. Together they took throughput to roughly 1.5 times the naive compile, and the same code now carries to Trainium2 1.

The agents did the porting work

Every optimization came out of a harness built on Flyte, a machine-learning operations platform 1. A candidate kernel is compiled, checked against the reference model to 0.999 cosine similarity, profiled, then run through a half-hour saturation benchmark on real hardware 1. Lightweight models grid-search compiler settings, reasoning agents test hypotheses and find bottlenecks, and the strongest code generators write the kernels that fix them 1. AWS's own Neuron Science team published a September playbook in which an agentic tool wrote a working Trainium kernel on its first attempt 6.

The cost side was the missing number

Exa's revenue is public: $100 million annualized, disclosed on October 13, on 2 billion search requests a month 2. Its May valuation prices the company at roughly 22 times that run rate 2.

The floor is public too: DeepSeek, the Chinese AI lab backed by the High-Flyer quant fund, charges 75 cents per 1,000 search requests, against $4 for Exa's cheapest tier 234. The cost side had no number until this week, and the only document for it is Exa's own post, tuned so far for Inferentia2 in Oregon 1.

The floor moved again on publication day: from November 18, DeepSeek stopped billing search calls made inside its own models separately and counts them as token usage instead 4.

DeepSeek's floor sits far beneath every listed search API

Exa's cheapest tier clears DeepSeek's floor five times over

$0/1k$2/1k$4/1k$6/1k$8/1kDeepSeek$0.75/1kfloor set October 17Exa Lite$4/1kBrave Search$5/1kPerplexity Sonar$5/1kExa Auto$7/1kTavily$8/1k
Data
Value
DeepSeek$0.75/1k
Exa Lite$4/1k
Brave Search$5/1k
Perplexity Sonar$5/1k
Exa Auto$7/1k
Tavily$8/1k
Pay-as-you-go list prices per 1,000 search requests as posted, October to November 2026; Perplexity adds token fees on top. Sources: The Inference, Exa and DeepSeek pricing pages.2,3,4

Amazon wins a reference workload, Nvidia does not

For Amazon, the post is a production inference workload from a company whose entire business is inference, running on its own silicon, and AWS already counts Anthropic$350B — Anthropic, private, latest valuation $350B, Databricks$218.01B — Databricks, private, latest valuation $218.01B and Uber−4.74% — Uber, down 4.74 percent today among Trainium customers 15. For Nvidia, the same workload is inference revenue it did not win 1.

Exa can cut prices only as fast as its cost per million documents falls. So the two things to watch are plain: whether Exa holds its $4 Lite price now that its unit-cost curve is public 3, and whether tuning done for one chip in one region carries to Trainium2 and other regions at the same gain 1.

Get Platforms by email, free

The cloud, data and developer software AI is built on. In our daily email, every weekday.

Loading AMZN

Reader comments

0 comments

    More in Platforms

    All in this topic →

    Platforms

    Cloudflare's CFO says expect a 3% beat, not 6%, as machine traffic hits 68%

    The stock rallied toward the street's $350 median target hours after the warning, inside a company whose insiders sold roughly $330 million of stock in six months without a single purchase. The risk of a smaller beat sits with holders, not with the AI labs that will end up paying for the web.

    Platforms

    Replit's CEO says revenue went from $3 million to $1 billion in two years, 9 times sales vs Cursor's 30

    The first top line Replit has ever put on the record makes it the cheapest of the big AI app builders per dollar of revenue: roughly nine times sales, against about thirty for Cursor and Lovable. The gap runs through token prices, and Microsoft, now Replit's distributor and model rival, reset those last week.

    Platforms

    Atlassian's index finds AI writes 70% of merged code, and the work left is what it sells

    The first quarterly AI SDLC Index shows code generation getting dramatically cheaper while the scarce work moved to human review, planning and documentation, the products Atlassian sells. It lands as investors ask whether AI deletes the company's per-seat business, and it runs on data from DX, a business Atlassian bought.

    Platforms

    Atlassian ships Secoda's data catalog for its AI agents; the test is holding ARR growth above 18.5%

    Ten months after quietly buying the data-catalog startup Secoda, Atlassian has wired data warehouses into the graph its agents read, at a price no filing discloses. The bet: whoever supplies agents their context gets paid whichever model wins, and investors who cut the stock 17 percent last week want the revenue proof.

    Platforms

    Dynatrace tells its largest customers to cut log costs in half while log consumption nears $250 million

    The advice comes from the vendor's own market chief, days after its CFO built an investor case on a log-consumption boom nearing $250 million. Until enterprises pay for the agent trace instead of the stored byte, the gap between those two stories sits with shareholders.

    Platforms

    Appian's first agent economics: an 80% token cut from giving agents less work

    The November 17 launch prices nothing, and the one named customer saved money by moving work off agents. That points demand at Appian's legacy process line, and APPN closed down 3.8%.

    Subscribe to TheInference for free