Send an email
For a while, “AI search” has been framed as a battle of models: bigger context windows, better retrieval, smarter agents. But the quieter battleground is more physical. It’s watts, bandwidth, and increasingly, memory
If you want a real-time indicator of where the industry’s constraints are tightening, look at RAM. Over the past few months, price spikes and availability issues have started showing up as broad supply-chain headlines. Reporting across the hardware ecosystem points to DRAM pricing pressure driven by AI infrastructure demand, with projections showing conventional DRAM contract price increases and continued strength into 2026
This matters for advertisers and publishers because AI search is, at its core, an ingestion problem. Before an AI system can “answer,” it must crawl, parse, dedupe, chunk, embed, and re-embed a planet’s worth of messy, unstructured web content. That pipeline is expensive in compute, expensive in energy, and expensive in memory.
RAM is the working set of the modern internet. If compute is the engine, memory is the fuel line. When memory gets tight, everything changes: what workloads get prioritized, what gets cached, what gets processed now versus later, and which inputs are considered “worth it.” And when memory gets expensive, the incentives shift from “index everything” to “index efficiently.”
We’re watching that shift happen in public, with steep year-over-year DRAM price increases tied to AI-led demand. The hardware press is also capturing second-order effects: GPU vendors citing memory constraints, and PC makers passing component costs through.
Now zoom out: AI search is becoming a competition for cost-to-understand. That’s the beginning of an energy war.
Another way to think about this is through marginal cost. Every additional document an AI system ingests carries a real resource burden: memory allocation, storage overhead, recomputation during model refreshes, and embedding updates as models evolve. At web scale, even small inefficiencies multiply into material infrastructure costs. What looks like harmless front-end clutter at the page level becomes measurable drag at the data center level.
That economic reality will shape discovery in subtle ways. When platforms must choose what to refresh, re-embed, or keep in high-speed memory tiers, they will favor sources that are stable, structured, and predictable. Efficiency becomes a ranking signal by proxy – not because anyone declared it one, but because physics and pricing quietly enforce it.
As LLM-powered discovery becomes mainstream, the systems that mediate attention will quietly prefer content that is cheaper to ingest. Not because they’re “biased” in the moral sense, but because they are budget-constrained in the physical sense. Parsing a chaotic page full of bloated scripts, shifting DOM elements, infinite scroll, cookie banners, and layout thrash costs more than reading a clean document. Extracting meaning from a PDF screenshot costs more than consuming structured text. Disambiguating entities in a sloppy article costs more than mapping a publisher-defined ontology.
Vectorization – representing meaning as embeddings that can be efficiently retrieved, compared, and composed – reduces cost. When content is consistently structured, semantically labeled, and paired with durable identifiers (entities, topics, products, locations), the ingestion pipeline becomes less speculative, with fewer retries and less deduping.
Less “figure it out later.” A smaller working set. Lower RAM pressure. Lower energy burn. The implication for publishers is blunt: your content strategy is becoming a systems strategy.
What does that look like in practice? It starts with making your best content machine-readable by default: Clean HTML that renders meaning without a JavaScript obstacle course. Stable page structures that don’t require a headless browser to interpret. Clear canonicalization so systems don’t embed five slightly different versions of the same story.
Then, go one step further: publish vector-ready artifacts. That can mean offering well-documented feeds and APIs for licensed partners. It can mean providing chunked article text with section headers preserved. It can mean exposing entity metadata (people, places, brands, tickers) in consistent schema. It can mean maintaining your own embedding layer for your archive so partners don’t have to repeatedly reprocess it.
Yes, this is technical. That’s the point. In the energy war of AI search, the winners will be the sources that lower friction and reduce compute waste. And the losers won’t necessarily be the least trustworthy – they’ll be the most compute-intensive.
If AI systems increasingly reward vector-native, machine-readable sources, then “format” becomes a moat. Media companies can differentiate with reporting and ingestibility: the ability to be reliably retrieved, attributed, and recomposed in AI answers without degrading meaning.
RAM price spikes are a reminder that the digital world still runs on scarce physical resources. When those resources tighten, ecosystems adapt. AI search will adapt by privileging efficiency. Publishers should adapt by designing content for vectors, not just visitors.
Because in the coming era, the question will be, “Is your content worth the watts?”
