Qualcomm's two new flagships share one spec sheet, and the memory clock that binds on-device AI has not moved

The one number on the sheet that did not move

Qualcomm's Snapdragon 8 Elite Gen 5 product brief, document 87-93124-1 Rev B, states the platform's memory support in two lines: "Support for LP-DDR5x memory, up to 5300MHz" and "Memory Density: Up to 24GB". The Snapdragon 8 Elite Gen 6 product brief, document 87-A3018-1 Rev A, published on 21 September 2026, states "Support for LP-DDR5x/6, up to 5300 MHz" and "Memory Density: Up to 24GB". The Snapdragon 8 Elite Extreme Gen 6 brief, 87-A3019-1 Rev A, states exactly the same two lines.

The clock is the same. The capacity ceiling is the same. On the published record, across a full generation, the memory subsystem that feeds every one of the AI blocks Qualcomm spent this launch talking about carries the same headline number it carried last year.

That matters more than it looks, because generating text with a language model is not a compute problem on a phone. It is a memory problem. To produce each token, the processor has to read the model's active weights out of DRAM. Compute units sit idle waiting for those bytes. The ceiling on how fast a local assistant can answer is memory bandwidth divided by the weight bytes read per token, and nothing on the NPU can raise a ceiling set by the bus.

Qualcomm announced both parts on 21 September 2026 at Snapdragon Summit. The press release names nine launch partners, HONOR, iQOO, Motorola, OnePlus, OPPO, REDMI, RedMagic, vivo and Xiaomi, and contains no numbers at all: no clock, no capacity, no percentage, no throughput. Every quantified claim in this briefing comes from the two product pages and the three product briefs, which are the only documents on the record that carry figures.

The blocks Qualcomm names, and the number missing from all of them

A silicon launch is a list of named blocks, and the names are the story, because each one is a claim about what the part does and a boundary around what the vendor is willing to be held to. The Gen 6 brief names eight functional blocks plus memory. Here is what each of them is, stated only from the brief.

Every named block in the Snapdragon 8 Elite Gen 6 product brief, 87-A3018-1 Rev A, with what the brief states and what it leaves out

Block	What the brief states	What the brief does not state
Oryon CPU	2 prime cores to 5.0 GHz, 6 performance cores to 4.0 GHz, 16 MB Oryon Flex Cache, 64-bit, 2nm	Which Oryon generation the base part uses. The base page says "next generation", the Extreme page says 4th generation
Adreno GPU	Matrix cores, hardware ray tracing with Lumen, Nanite, Vulkan 1.4, OpenCL 3.0	Any clock, any core count, and the size of Adreno high-performance memory on the base part
Hexagon NPU	12 scalar, 8 vector, 1 tensor, 1 element; INT2 to FP16 with mixed precision; Micro Tile Inferencing; 64-bit memory virtualisation	Throughput in TOPS at any precision, the NPU clock, and the size of NPU shared memory
Sensing Hub	Dual micro NPUs for audio, voice and sensors; dual always-sensing ISPs for two concurrent always-sensing cameras; Personal Scribe	Power draw, what the personal knowledge graph retains, for how long, and where it is stored
Spectra ISP	Triple 20-bit AI-ISPs, 8K capture at 30 FPS, up to 320MP photo capture, 16 bits of AI data per pixel	Whether always-sensing capture is visible to the host operating system or to a management agent
X105 5G Modem-RF	14.8 Gbps downlink, 4.2 Gbps uplink, 6 receive antennas, 6xCA at 500 MHz, 3GPP Release 19 hardware, NR-NTN and NB-NTN	Which satellite operators or networks the non-terrestrial modes actually work on, and when
FastConnect 8800	11.6 Gbps peak, 802.11bn, 4x4 Wi-Fi, Bluetooth 6.0 with channel sounding, UWB, Thread 1.4	Any Wi-Fi security modes at all. The Gen 5 brief enumerated four WPA3 modes; the Gen 6 briefs enumerate none
Security	Trust Management Engine as Root of Trust, Qualcomm TEE, Type-1 Hypervisor, Secure Processing Unit, wireless edge services, 3D Sonic, secure C2PA	Any certification, evaluation scheme, assurance level or attestation interface for any of those six items
Memory	LP-DDR5X or LP-DDR6, up to 5300 MHz, up to 24 GB density	Bus width, channel count, or any bandwidth figure in gigabytes per second

What the memory figure actually permits, with the arithmetic shown

Qualcomm does not publish a bandwidth figure, so here is the calculation and its assumption, stated plainly so it can be argued with.

The brief states 5300 MHz. LPDDR is double data rate, so the transfer rate is 5300 x 2, or 10,600 mega-transfers per second. Bandwidth is transfer rate multiplied by bus width. The bus width is the number Qualcomm does not give. Eight-series Snapdragon parts have conventionally used a 64-bit memory interface, so taking 64 bits, or 8 bytes, per transfer:

10,600 MT/s x 8 bytes = 84,800 MB/s, which is 84.8 GB/s.

That figure is inference, not reporting. If the interface is wider, or if an LPDDR6 configuration organises its channels differently, the real number is higher and Qualcomm has chosen not to say so.

Take the 84.8 GB/s and divide by the weight bytes a dense model must read per token. An 8 billion parameter model quantised to 4 bits holds 4 GB of weights, so 84.8 divided by 4 gives a ceiling of 21.2 tokens a second. The same model at 8 bits holds 8 GB and halves to 10.6. A 13 billion parameter model at 4 bits is 6.5 GB and gives 13.0. A 30 billion parameter dense model at 4 bits is 15 GB of weights and gives 5.7, which is slower than most people read.

These are ceilings, not benchmarks. Real throughput is lower, because attention state and activations compete for the same bus and no memory system runs at its theoretical peak. The shape of the curve is the point.

Capacity is the second wall, and it is the harder one. The 24 GB ceiling is the whole platform's memory, shared with Android, the launcher, the browser and every background app. Nothing like all of it is available for model weights. A 30 billion parameter dense model at 4 bits needs 15 GB resident before a single token of context is cached, which on a 24 GB handset is not a realistic steady state and on a 12 GB handset is impossible.

This is precisely where mixture-of-experts architectures earn their keep, because they decouple the two walls. A sparse model holds all its weights in memory but reads only the active experts per token, so it pays the capacity cost in full and only a fraction of the bandwidth cost. Launch coverage reports Qualcomm claiming the Extreme part's NPU can run mixture-of-experts models above 30 billion parameters. That claim appears in neither product brief and on neither product page. It rests on secondary reporting of a Qualcomm statement, and this briefing treats it as unverified.

Two parts, one spec sheet, and a difference nobody has quantified

Qualcomm's own framing is a "multi-flagship strategy" giving OEMs choice. The press release says the base part is "built on the same advanced 2nm process node" and powers "many of the same breakthrough AI, gaming, imaging, and connectivity features" as the Extreme.

Read the two specs and features sheets side by side and the word "many" is doing very little work. The CPU section is identical, character for character: two prime cores to 5.0 GHz, six performance cores to 4.0 GHz, 16 MB of Oryon Flex Cache, 64-bit. The memory section is identical. The display, audio, location, charging, modem and Wi-Fi sections are identical. The security section is identical. The part numbers differ, SM8950 against SM8975.

Where the two Gen 6 parts actually diverge on the record

Compiled line by line from the Snapdragon 8 Elite Gen 6 Product Brief (87-A3018-1 Rev A) and the Snapdragon 8 Elite Extreme Gen 6 Product Brief (87-A3019-1 Rev A)

Item	8 Elite Gen 6	8 Elite Extreme Gen 6	Not stated for either
CPU	2 prime to 5.0 GHz, 6 performance to 4.0 GHz, 16 MB Flex Cache	Identical, character for character	Why identical silicon yields 10 per cent on one part and 13 per cent on the other
Process and memory	2nm; LP-DDR5X/6 to 5300 MHz; 24 GB; UFS 5.0	Identical	Foundry, bus width, or any bandwidth figure
NPU	12 scalar, 8 vector, 1 tensor, 1 element; Element Accelerator; INT2 to FP16	Same, plus one extra line: 50 per cent larger shared memory	The size of that shared memory, before or after the 50 per cent
GPU	Adreno GPU with matrix cores; ray tracing; Nanite	Adreno Neural Fusion; matrix cores; MegaLights; 18 MB high-performance memory	Core count or clock for either part
Video decode	H.265, VP9, AV1	Adds H.266	Whether the base part can gain H.266 in software
Camera	8K at 30 FPS; 4K slow motion at 120 FPS	8K at 60 FPS; 4K slow motion at 240 FPS; APV capture	Which sensors and thermal envelopes those modes require
Security block	Seven lines	The same seven lines, word for word	Any difference at all between the two parts
Part number	SM8950	SM8975	Whether these are the same die at different bins

So the honest answer to "what is different" is: the GPU, one NPU memory line, one video decoder, and the camera pipeline. Not clocks, not core counts, not cache, not memory support. Every published CPU and memory figure is shared.

Which makes the uplift claims hard to read. The base part claims 10 per cent better CPU performance against Gen 5 and the Extreme claims 13 per cent, against the same baseline, from a CPU section that is identical on both sheets. Qualcomm names no workload for either figure. There are perfectly ordinary explanations, sustained thermal headroom, memory controller tuning, a different bin, a different benchmark, and Qualcomm states none of them. A reader cannot tell from the record which one applies, and that is a fact about the record rather than about the silicon.

The security block is a list of names, not a level of assurance

Here is the whole of what Qualcomm claims about security on both Gen 6 parts, quoted in full because it is short: a Trust Management Engine described as a Root of Trust "with platform-level security foundations"; a Qualcomm Trusted Execution Environment; a Qualcomm Type-1 Hypervisor "for isolation from the high-level OS"; a Secure Processing Unit; Qualcomm wireless edge services; the 3D Sonic and 3D Sonic Max fingerprint sensors; "Secure C2PA AI provenance data support"; and "Enhanced AI processing isolation and model protection".

That is the complete list. It is identical on the base part and the Extreme.

Every item in it is a name for a place where code runs or a thing that exists. Not one is a statement about what has been evaluated, by whom, against what, or what an attacker would have to do to get past it. "Trusted Execution Environment" describes an architecture, not an assurance level. A TEE is a boundary; whether that boundary holds depends on the firmware inside it, the interfaces exposed to it, and the update path that keeps it patched, and none of those are in scope for a product brief. "Enhanced AI processing isolation" is enhanced relative to something the brief does not identify. "Secure C2PA AI provenance data support" is the only line that names an external standard and is therefore the only line a procurement team can test against a published specification.

This is not a criticism of the engineering, which may well be excellent. It is a statement about what the document does. A product brief is a sales artefact aimed at handset makers, not a security target, and it should be read as one.

What is more interesting is the comparison with last year's sheet, because four qualifications present in the Gen 5 brief are not in either Gen 6 brief.

Security and Wi-Fi lines quoted from the Snapdragon 8 Elite Gen 5 Product Brief (87-93124-1 Rev B) against the corresponding lines in both Gen 6 briefs

Line	Snapdragon 8 Elite Gen 5 brief	Both Gen 6 briefs
Trusted Execution Environment	"Qualcomm Trusted Execution Environment (TEE) and Qualcomm TEE Services for use cases requiring higher processing assurance"	"Qualcomm Trusted Execution Environment (TEE)"
Secure Processing Unit	"Secure Processing Unit (SPU) with support for the Android latest Strongbox software components"	"Secure Processing Unit (SPU)"
Wireless edge services	"Qualcomm wireless edge services (WES) for secure attestation and provisioning"	"Qualcomm wireless edge services (WES)"
NPU model protection	"Support for Gen AI model encryption for extra security", listed under the NPU	No equivalent line under the NPU on either part
Wi-Fi security	"Wi-Fi Security: WPA3-Enterprise, WPA3-Enhanced Open, WPA3 Easy Connect, WPA3-Personal"	No Wi-Fi security line of any kind

What on-device inference does to a BYOD or MDM policy

Strip the marketing away and the governance consequence of this launch is one sentence: the pitch is an assistant that reads your device, learns from your behaviour and acts across your apps, and does it all without the data leaving the handset.

Qualcomm says this in its own words. The base product page describes "personalized on-device agents that anticipate and act on your behalf" with "on-device learning adapting to your preferences over time keeping your data safe and secure". The base brief describes the Sensing Hub, the Hexagon NPU and the Oryon CPU working together so the platform can "understand your intent, access your personal context, map out a plan, and coordinate tasks across apps". The Extreme page promises "a persistent understanding of your world".

For a security team that is a genuinely new shape of problem, and the usual controls do not address it. Nearly every mobile control in common use is built around the assumption that something interesting leaves the device: a proxy sees the request, a cloud access broker sees the application, a data loss prevention rule sees the payload, a SIEM records the event. Local inference breaks all four at once. If the model runs on the NPU, there is no request, no payload and no event. The prompt containing a customer name, a contract clause or an unannounced financial figure is composed, processed and answered inside a package you cannot instrument and your MDM cannot see.

This is not a reason to panic, and it is arguably better for privacy than sending the same text to a cloud endpoint. But it is the end of egress as the place where mobile data governance happens, and no equivalent control has replaced it.

Three specific consequences follow from the spec sheet rather than from speculation.

Always-sensing is a design feature, not an edge case. Both briefs state that the Sensing Hub carries "Dual Always-Sensing ISPs to support two concurrent always-sensing cameras" and dual micro NPUs "for creating personal knowledge graphs". A persistent behavioural corpus built on the handset is the explicit product. Where that corpus lives on a dual-persona device, whether inside the Android work profile or outside it, is not stated in any of these documents, and it is an OEM and operating system question rather than a silicon one.

On-device learning breaks reproducibility. The Extreme page promises "on-device AI learning", and the base brief says the model adapts to "how you edit photos, write messages, and personalize visual experiences". If the model on the device changes with use, two handsets on the same corporate build will not behave the same way, and a behaviour you observed in testing may not be the behaviour in the field. Any assurance argument that relies on a fixed, testable configuration needs rewriting.

Provenance is the one control that produces readable evidence. "Secure C2PA AI provenance data support" is listed on both parts. C2PA is a published standard with verifiable signed manifests, which means media captured or edited on these devices can in principle carry an attestable record of how it was made. For anyone whose exposure is synthetic media rather than data leakage, that line is the most useful sentence in the whole brief.

What to do, in the order worth doing it

The question that exposes the gap

Qualcomm has built a launch around agentic AI running locally, on a platform whose published memory clock and memory ceiling are identical to last year's, with no throughput figure for the processor doing the work and no bandwidth figure for the bus that feeds it. The two parts share a CPU section, a memory section and a security section, and differ by 3 percentage points on a CPU claim neither sheet explains.

None of that means the silicon is not faster. It very probably is. It means the published record cannot tell you by how much, on what, or within what limits, and a buyer committing a fleet to on-device AI is being asked to take the direction of travel on trust.

So the question to put to the handset maker, and to put in writing:

If the answer to the last part is "nothing", that is the honest state of mobile data governance in 2026, and it is worth knowing before the fleet arrives rather than after.

Key facts