The cheap part got cheaper
AWS just cut the query price on S3 Vectors by up to 80% for indexes over 10 million vectors. Automatic, every region, no application changes. Ethan Steininger of Mixpeek flagged it as the move almost nobody noticed, and read it the way a lot of people will: that's the tier where standalone vector databases earn their margin, and AWS just repriced it. His conclusion: "the dedicated vector database isn't dead, but it's getting compressed into a feature. The database was never the product. Retrieval quality is."
I think he's right twice over, which is what makes the reading odd. If a vector index really is just a data structure, and retrieval quality really is the product, then the price cut he's reading as a turning point is closer to a non-event. It touches the one slice of retrieval that value was never in. Nothing got compressed here except the floor everyone was already standing on.
The concession
A vector index is a data structure, not a database. I don't mean that as a slight. An approximate-nearest-neighbour index is a graph or a set of quantised centroids that answers one question fast: given this point, which stored points are closest? Put it on object storage, put it in RAM, put it wherever - the operation is the same. It's a lookup.
And a lookup should get cheap. Commodity single-vector retrieval - one embedding per document, cosine similarity, top-k - is the most standardised thing in the whole stack. Everyone does it the same way, the algorithms are public, and the hard part is engineering, not insight. When something is that uniform, the hyperscalers absorb it and drive the price toward the cost of the disk it sits on. S3 Vectors is built for exactly that: object-storage-backed, aimed at the cold tier - vectors you query occasionally and don't need back in single-digit milliseconds - and priced accordingly. It getting 80% cheaper at scale is the natural order playing out. No argument.
And cheap is not nothing. Single-vector retrieval that costs almost nothing to run at the scale this repricing targets unlocks things - RAG over enormous, rarely-touched corpora; archives you'd never have paid to keep queryable. For a large share of real workloads, one averaged vector per document is entirely good enough, and making that tier cheaper expands what's economic to build. That's a real win on cost, which is a different axis from differentiation.
The turn
But hold the two claims next to each other. "The cheap lookup got cheaper" and "retrieval quality is the product" are pointing in opposite directions.
Retrieval quality isn't moving down the compute curve. It's moving up it. The techniques that actually change what comes back - late interaction, multi-vector representations, metadata-aware hybrid scoring - all spend more compute per query, not less. Late interaction keeps a vector per token instead of averaging a document down to a single point, then scores query tokens against document tokens at retrieval time. That's more storage, more math, more work at query time. It is the deliberate opposite of the thing AWS just made cheaper.
I've written before about why the single averaged vector fails on compound, high-intent queries - it lands near two requirements and on neither, because you can't be extreme on two independent axes and still sit near the query when everything you know about a document has been squeezed into one point. That failure is a quality failure, and no price cut on the single-vector lookup touches it. The 80% discount lands squarely on the slice of retrieval where quality is least differentiated - the slice where the mechanism is identical for everyone and nobody is pulling ahead.
What actually got compressed
So what's being compressed into a feature is the commodity slice - the raw ANN lookup that was already the least defensible thing anyone was selling. That slice was always going to zero. Good.
What's left is getting bigger and harder, not smaller. Filtering that stays correct at scale. Freshness. Hybrid scoring that blends lexical and semantic signal without one drowning the other. Multi-vector and late-interaction retrieval that keep a document's facets intact instead of averaging them away. None of that is a lookup. None of it gets solved by cheaper object storage. And all of it is where the difference between a demo and a system that a buyer trusts actually lives.
This is the same shape I keep running into. The value moves up the stack, the components underneath commoditise, and people mistake the falling price of the commodity for the falling value of the whole. The index becoming a feature doesn't compress the database - it clears the floor so the interesting work is easier to see.
Taking his last line further
"The database was never the product. Retrieval quality is." I'd sign that. I'd just take it one step further than the pricing news does.
If retrieval quality is the product, then a cheaper nearest-neighbour lookup is a line-item win and a strategic non-event. It makes the part that was never your moat a bit less expensive to run. Worth having - nobody turns down 80% off. But if you were counting on the raw index as the thing that set you apart, the price cut didn't compress your moat. It just showed you that you never had one there.
The interesting work was never in the index. AWS repricing the index is the clearest evidence yet of exactly that.