Did Google’s TurboQuant Actually Solve AI Memory Crunch?

On March 25, 2026, Google Research revealed a weblog submit a couple of compression algorithm referred to as TurboQuant.

Cloudflare’s CEO Matthew Prince referred to as it “Google’s DeepSeek moment.” Half the web in contrast it to Pied Piper, the fictional startup from HBO’s Silicon Valley. The memes moved sooner than the precise analysis.

So what truly occurred? And does this algorithm change something in regards to the reminiscence state of affairs the AI trade has been panicking about for the previous 18 months?

Let’s decode.

Why Modern AI Is So Hungry for Memory

When an LLM generates textual content, it does not recompute all the things from the start with each new phrase. Instead, it shops all its prior calculations in a fast-access buffer referred to as the key-value cache, or KV cache. Every token the mannequin has seen in a dialog will get saved there, so when the mannequin processes the following token, it will possibly look again at what got here earlier than with out redoing all the mathematics.

The downside is the cache grows repeatedly. A mannequin working via a 100,000-token doc is holding an enormous quantity of energetic information in GPU reminiscence simply to take care of context. And this acquired considerably worse when reasoning fashions grew to become mainstream. Reasoning means lengthy context, lengthy context means a big KV cache, massive KV cache means you want loads of reminiscence. By 2024, anybody taking note of the trajectory of AI fashions might see the place this was heading and the market principally did not catch up till costs began reflecting it.

And the trade has been preventing this downside for years, with real ingenuity, and TurboQuant is the newest step in that arc.

What TurboQuant Is and How It Works

TurboQuant compresses that KV cache down to three bits per worth, from the usual 16. The claimed discount is 6x in reminiscence footprint, with an 8x speedup in consideration computation on Nvidia H100 GPUs, and no measurable accuracy loss in benchmarks.

The math works in two levels. 

The first stage, PolarQuant, converts information vectors from Cartesian coordinates into polar coordinates. In Cartesian type, a degree is described by how far it sits alongside the X axis and Y axis: a grid of (x, y). In polar type, the identical level is described by its distance from the origin (r) and the angle it makes from a reference route (θ). The conversion is: r = √(x² + y²) and θ = arctan(y/x). Going again: x = r·cos(θ) and y = r·sin(θ). In greater dimensions, the identical precept extends.

Why this issues for compression is as a result of in polar area, the angular distribution of AI consideration information clusters in predictable, concentrated patterns. Traditional quantization strategies need to retailer further normalization constants alongside compressed information so the system can decompress precisely later. Those constants add one or two bits per worth proper again in, partially undoing the financial savings. PolarQuant eliminates that overhead as a result of the construction of the information in polar area makes these constants pointless.

The second stage handles the residual error left over from stage one. Each leftover error quantity will get lowered to a single signal bit, constructive or unfavourable. That signal bit acts as a statistical zero-bias corrector, which means the compressed cache stays equal to the full-precision unique when the mannequin computes consideration scores. The mannequin does not discover the distinction.

Google examined TurboQuant on 5 commonplace benchmarks for long-context fashions, together with LongBench and Needle in a Haystack, utilizing Gemma, Mistral, and Llama. At 3 bits, it matched or beat KIVI, the usual baseline for KV cache quantization. On needle-in-a-haystack duties the place the mannequin has to find a particular reality buried in a protracted doc, it hit excellent scores at 6x compression. 

The Crunch That Was Years within the Making

In 2023, reminiscence producers have been shedding cash. DRAM costs had collapsed after the pandemic oversupply, and Samsung, SK Hynix, and Micron all pulled again on capital expenditure. They weren’t constructing new fabs as a result of there was no margin to justify it. But it coincided exactly with the start of the reasoning mannequin period, which was about to create a requirement curve nobody had seen earlier than on this trade.

Let’s perceive why AI is so exhausting on reminiscence. A GPU wants information to maneuver at excessive speeds to maintain its processors fed. An HBM4 stack, the kind of reminiscence utilized in Nvidia’s newest chips, transfers reminiscence at roughly 2.5 terabytes per second. A comparable space of ordinary DDR5, the reminiscence in your laptop computer, does someplace round 64 to 128 gigabytes per second. Consumer reminiscence is constructed for a very totally different job. 

HBM is constructed otherwise, stacked in a number of layers, related with 1000’s of micro-connections referred to as through-silicon vias, and it is terribly costly to provide. Producing one gigabyte of HBM consumes 4 instances the wafer capability of ordinary DRAM. To put that in GPU phrases: a single Nvidia H100 presently prices between $25,000 and $30,000 per chip, and reminiscence accounts for roughly 30% of the price of deploying AI at scale. When Meta constructed its preliminary H100 coaching cluster with 24,000 of these chips, the GPU {hardware} invoice alone crossed $800 million, earlier than a single energy cable was run or a server rack assembled. That’s one cluster, hyperscalers are constructing dozens. Of the $600 billion in mixed Big Tech capital spending this 12 months, roughly $180 billion goes to reminiscence alone.

People normally make the “just make more memory” argument. Global silicon wafer manufacturing capability is rising, however solely at round 6 to 7% per 12 months. AI infrastructure spending is rising at charges many instances that. The fabs that can finally shut the hole began building after the demand sign hit, which implies the significant new capacities do not come on-line till 2027-2028 and the crunch can probably final till 2030.

The Compression Arms Race That Was Already Happening

The trade has been chipping away on the KV cache reminiscence downside for years.GPT-2 XL, the biggest 2019 variant, used the only attainable design: each consideration head stored its personal impartial set of keys and values. Cost: round 300 kilobytes per token. By 2024, Llama 3 8B launched grouped-query consideration, the place a number of heads share the identical saved representations as a substitute of sustaining separate copies. Cost dropped to 128 kilobytes per token, lower than half, with virtually no high quality loss on benchmarks. Then DeepSeek V3 went additional with multi-head latent consideration, compressing the key-value pairs right into a lower-dimensional type earlier than storing them and decompressing at inference time. Cost: 68.6 kilobytes per token, on a mannequin with 671 billion complete parameters, although solely 37 billion are energetic at any second.

That development, 300 to 128 to 68 kilobytes per token, is the compression arc that existed earlier than TurboQuant confirmed up. Each step traded one thing, normally some architectural complexity or slight recall degradation, for significant reminiscence financial savings. Each step additionally captured the better positive aspects first. What remained acquired more durable.

So by the point TurboQuant arrived, the low-hanging fruit was gone. TurboQuant issues much less as a result of it saves further reminiscence and extra as a result of it marks the place KV cache compression is approaching the information-theoretic restrict. You’re near the Shannon ceiling. Every further bit squeezed out from right here prices extra engineering effort and dangers extra high quality degradation than the final.

There’s additionally an issue no compression algorithm touches. When the KV cache grows too massive for obtainable GPU reminiscence, fashions typically summarize their very own context right into a shorter type and proceed from the abstract. The compression is lossy in methods the mannequin cannot detect. A selected price range determine turns into “approximately that amount.” A nuanced instruction turns into “something about guidelines.” The mannequin retains going, assured in data that now not absolutely exists. Compression makes the cache smaller. It does not clear up the issue of deciding what’s truly price retaining.

So Why the Market Reaction Was Wrong

TurboQuant solely addresses inference reminiscence, particularly the KV cache throughout inference. Training a mannequin, the months-long, multi-billion-dollar strategy of educating the mannequin within the first place, requires essentially totally different reminiscence, pushed by activations, gradients, and optimizer states. TurboQuant has zero impact on any of that. The large HBM buildout that hyperscalers are funding exists primarily to coach and retrain ever-larger fashions. That demand curve is untouched by a KV cache compression algorithm.

Beyond coaching, TurboQuant is a analysis outcome with no manufacturing deployment. The paper was initially revealed in 2025 and acquired re-featured on the weblog forward of ICLR. Google itself hasn’t deployed it extensively within the 12 months because the math was first documented.

The market has now seen this precise film twice, however panicked each instances. Weird proper?

So What TurboQuant Actually Changes

The algorithm does have actual implications. They’re simply totally different from what the market priced in.

The most instant is inference economics. TurboQuant compresses the KV cache, which determines what number of concurrent customers a single GPU can serve and the way lengthy a context window is sensible at scale. If it will get deployed throughout manufacturing inference stacks, the throughput per GPU will increase. That issues for AI merchandise operating thousands and thousands of queries per day, the place inference value is the recurring expense that determines profitability. Anything that modifications the memory-to-compute ratio per question shifts the fee construction of operating AI merchandise.

The longer-term implication is on-device AI. Right now, operating a succesful language mannequin domestically on a cellphone or laptop computer requires both compromising on high quality or shopping for costly {hardware}. If TurboQuant’s method will get carried out in native inference runtimes at scale, the {hardware} ground for operating a significant AI mannequin drops. Models that presently require cloud infrastructure might run domestically.  But it performs out over years, not quarters, and it has extra to do with software program ecosystem adoption than with whether or not reminiscence chip shares are accurately priced as we speak.

It’s positively actual math that compresses one particular sort of reminiscence utilization throughout one section of AI operation. But it does not construct fabs and it does not change coaching economics. Memory will get in-built clear rooms in South Korea and Idaho, by individuals working instruments that value a whole bunch of thousands and thousands of {dollars} every. That a part of the provision chain strikes on a very totally different clock than an algorithm (or only a analysis paper.)

So the crunch solely ends when the fabs are finished.