Google's turboquant slams memory stocks, but the sell-off smells like panic
Memory-chip investors just watched $12 billion evaporate in Seoul and Tokyo because Google claims it can run large language models with 83% less RAM. The math looks brutal: SK Hynix down 6.4%, Kioxia ditto, Micron already bruised from Wednesday’s New York session. One paper from Mountain View and the whole memory complex wobbles.
The algorithm that shrank dram
TurboQuant is not another academic footnote. It is a production-ready compression layer that keeps model weights in a 4-bit representation yet recovers accuracy on the fly. Google’s internal benchmarks show inference latency dropping 8× while cutting working-set size by a factor of six. For hyperscalers that rent Nvidia H100 nodes by the millisecond, that translates into real margin.
Wall Street’s first reflex: if servers need fewer chips, memory makers are toast. Hence the knee-jerk dump of anything that smells like DRAM or NAND. But the Street is trading a 19th-century coal thesis against 21st-century cloud economics.
Jevons rides again
William Stanley Jevons observed in 1865 that more-efficient steam engines increased coal consumption, because they made energy cheap enough to invent new factories, railways, entire cities. JPMorgan’s trading desk dusted off the same paragraph Thursday morning: cheaper inference begets more inference. Morgan Stanley’s Shawn Kim spells it out in his note to clients: “Lower per-unit cost expands the TAM faster than it cannibalizes unit demand.”
Memory prices have tripled since early 2023. Spot 128 Gb DDR5 quotes at $3.90, up from $1.20 eighteen months ago. Fabs are running flat-out; Hynix’s Q1 capex is already spoken for by Nvidia. Even if Google hands every cloud provider a 6× memory coupon, supply is still capped by EUV hours and neon gas lines. Ortus Advisors’ Andrew Jackson puts it bluntly: “There is no idle capacity to destroy.”
What actually breaks
The immediate casualty is valuation, not volume. Kioxia’s 700% rally since August was begging for an excuse to cool. Same for the memory-tracking ETFs that doubled their AUM in four months. TurboQuant gives portfolio managers a narrative to lock in gains before the half-year mark.
Long term, the tech tilts power toward whoever owns the model stack. If a hyperscaler can serve the same tokens with one-sixth the DRAM, that saving flows straight to operating income. Memory vendors become mix-shifted, not eliminated: high-bandwidth HBM still ships, commodity DDR5 ships less. Margins compress at the bottom, widen at the top.
Google wins twice. It cuts its own inferencing bill, and it rents a cheaper stack to outsiders via Vertex. Amazon’s rumored purchase of humanoid-robot startup Fauna Robotics the same day only underlines the theme: hyperscalers are hunting any lever that makes compute cheaper, hungrier, more ubiquitous.
The sell-off will reverse the moment first-quarter earnings show bit demand still outpacing wafer starts. When that happens, remember the snapshot: traders panicked over a compression paper while fabs remained sold out through Christmas. Jevons, as always, collected the dividend.