Product 1 publisher3 min readPublished September 10, 2026 at 4:52 PM PDT

Positron bets $875m that smartphone memory can carry inference

Positron is now worth $5 billion on a design that swaps HBM for the memory that ships in phones, but Asimov has not taped out, mass production is set for the second half of 2027, and the 26x tokens-per-dollar figure comes from simulation.

The Product Desk · Product desk

Illustration accompanying Positron bets $875m that smartphone memory can carry inference

What happened

The appliances use LPDDR5X, the memory variety that mostly goes into smartphones, which is cheaper and easier to obtain than HBM but carries significantly less bandwidth.
Neither Titan nor the Asimov chip is in production, and the claim of up to 26 times the tokens per dollar of Nvidia's Blackwell GB300 NVL72 comes from Positron's simulations.
The company expects to tape out Asimov at the end of the year on TSMC's three-nanometer node, with mass production following in the second half of 2027.

Compiled by The Product DeskSomething wrong?How this is made

Why it matters

constraint No capacity plan written this year gets out of the HBM queue through Positron, because the part that would do it has not been fabricated and cannot ship in volume before the second half of 2027.
cost Whoever commits early absorbs the gap between a simulated rack result and measured throughput on first silicon, and that ratio is the entire price argument for buying the appliance at all.
exposure Positron has taken on a second supply dependency in place of HBM: its ramp rests on LPDDR5X commitments it says it still has to secure from partners.

The utilization claim is the load-bearing part of this design, and it can be sized. Positron says most AI accelerators use less than 30% of their HBM modules' bandwidth, while its own appliances unlock more than 90% of LPDDR5X's throughput ^[6]. That is a factor of at least three in bandwidth delivered per unit of peak bandwidth bought ^[2]. The substitution works if LPDDR5X's peak deficit against HBM is smaller than three. Positron published the two utilization percentages and left out the peak-bandwidth comparison against a server card ^[6].

Titan's specs can at least be checked against each other. Up to 18.4 terabytes of LPDDR5X delivering 23.68 terabits per second ^[7] is about 2.96 terabytes per second ^[3], so reading the entire pool once takes roughly 6.2 seconds ^[4]. No usable token rate survives a full sweep per token, which means the hardware fetches a subset of the stored weights on each pass ^[6]. That is what the activation-function modules are for, and Positron says they determine which of a model's neural networks participate in an inference task ^[8].

About ten months separate this month's announcement from the earliest mass production date the company has given ^[5]. The tape-out has to land first, at the end of the year, on TSMC's three-nanometer node ^[13]. "Our focus now is to tape out Asimov, bring Titan to production, and scale manufacturing to meet the demand in front of us," said Chief Executive Officer Mitesh Agrawal ^[11]. "This financing gives us the resources to do exactly that." ^[12]

Five times February works back to roughly $1 billion for the company seven months ago ^[1]. NEA, Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital and Silicon Graphics and Netscape co-founder Jim Clark led the round, alongside more than a dozen institutional investors ^[2]. Positron says its manufacturing ramp will emphasise securing supply commitments from partners for LPDDR5X ^[13], the part it is designing around and the one that mostly goes into phones ^[5].

Who this is for turns on where the ceiling sits on a given inference workload: memory capacity, or arithmetic throughput on a model that already fits in the boxes on the floor. It also turns on whether the install date can move to 2028. A buyer in the memory-bound, patient corner is the one Titan is built for, and Positron's pitch there is a single appliance running a 32-trillion-parameter model with a 10-billion-token context window ^[7]. A buyer whose models fit and whose problem is tokens per second this quarter is still queueing for HBM, and the interconnects that attach it are short too ^[4].

Positron's 26x is a rack-level simulation result, and the silicon it describes has not been fabricated ^[10]. Teams under capacity pressure tend to treat a tokens-per-dollar ratio as a procurement fact.

What to watch

Whether Asimov actually tapes out on TSMC's three-nanometer node by the end of the year, since the 2027 production date is timed off it.
Measured tokens per dollar on real silicon, to test the simulated 26x against the GB300 NVL72.
A first named Titan customer, and whether anyone orders at the 16,384-accelerator cluster scale Positron advertises.

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Positron AI Inc., a developer of inference appliances that do not include HBM memory, announced that it has raised $875 million in funding.
ReportedSupported2 sources— sign in to open View cited source
[2]
The round was led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital and Silicon Graphics Inc. and Netscape Inc. co-founder Jim Clark, joined by more than a dozen institutional investors.
ReportedSupported2 sources— sign in to open View cited source
[3]
Positron is now worth $5 billion, a fivefold increase from February.
ReportedSupported2 sources— sign in to open View cited source
[15]
The report of the funding round was published on September 10, 2026, and says Positron announced the raise that day.
ReportedSupported2 sources— sign in to open View cited source
[4]
The supply of HBM is currently well below demand, and there is also a shortage of the interconnects used to integrate HBM memory with graphics cards' computing circuits.
ReportedSupportedView cited source
[5]
Positron's inference appliances substitute HBM with LPDDR5X, a memory variety mainly used in smartphones; LPDDR5X is more readily available than HBM and costs less, but has significantly less memory bandwidth, which usually translates into slower inference.
ReportedSupportedView cited source
[8]
Titan uses a custom chip called Asimov, built around a systolic array whose modules include co-located memory holding LLM weights; the array is supported by modules optimized to run activation functions, which determine which of an LLM's neural networks should participate in an inference task, plus CPU cores that function as a "programmable escape hatch".
ReportedSupportedView cited source
[9]
Each Titan appliance includes up to eight Asimov chips, and customers can link multiple systems into clusters with up to 16,384 accelerators.
ReportedSupportedView cited source
[10]
Titan and Asimov are not yet in production; according to Positron, simulation data indicates a server rack powered by its silicon can process up to 26 times as many tokens per dollar as Nvidia's Blackwell GB300 NVL72 appliance.
ReportedSupportedView cited source
[11]
"Our focus now is to tape out Asimov, bring Titan to production, and scale manufacturing to meet the demand in front of us," said Positron Chief Executive Officer Mitesh Agrawal.
ReportedSupportedView cited source
[12]
"This financing gives us the resources to do exactly that."
ReportedSupportedView cited source
[13]
Positron expects to tape out Asimov at the end of the year using TSMC's three-nanometer node, with mass production set to follow in the second half of 2027, and the manufacturing ramp will place an emphasis on securing LPDDR5X supply commitments from partners.
ReportedSupportedView cited source
[14]
Reno, Nevada-based Positron says it has sidestepped the supply chain crunch.
ReportedSupportedView cited source
[d1]
A $5 billion valuation that is five times the February mark implies Positron was worth about $1 billion in February.
DerivedSupported1 source— sign in to open
[d3]
23.68 terabits per second of memory bandwidth is about 2.96 terabytes per second.
DerivedSupported1 source— sign in to open
[d4]
Reading Titan's full 18.4 terabyte memory pool once at 2.96 terabytes per second takes about 6.2 seconds.
DerivedSupported1 source— sign in to open
[d5]
From the September 10, 2026 announcement to the start of the second half of 2027 is about ten months.
DerivedSupported1 source— sign in to open
[d6]
Because a full read of Titan's memory pool takes about 6.2 seconds, per-token inference on the appliance must read only a subset of the stored weights.
DerivedSupported1 source— sign in to open
[6]
According to Positron, most AI accelerators use less than 30% of their HBM modules' memory bandwidth, while Positron's inference appliances unlock more than 90% of LPDDR5X's throughput; the company says the increased hardware utilization makes up for the difference in theoretical peak performance.
ReportedInsufficientView cited source
[7]
Positron's flagship system, Titan, ships with up to 18.4 terabytes of LPDDR5X providing 23.68 terabits per second of memory bandwidth, which Positron says lets a single Titan appliance run an LLM with 32 trillion parameters and a context window of 10 billion tokens.
ReportedInsufficientView cited source
[d2]
Positron's utilization figures imply at least a threefold advantage in bandwidth delivered per unit of peak bandwidth, so the LPDDR5X substitution only holds if its peak bandwidth deficit against HBM is smaller than three.
DerivedInsufficient1 source— sign in to open

Sources

Publishers with included, body-backed reporting in this cluster.

siliconangle.com
2 articles · September 10, 2026
Chipmaker Positron nabs $875M to speed up inference with consumer-grade memory

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Try

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

The claim a builder can use here is the utilization one. Positron told SiliconANGLE that most accelerators run under 30% of their HBM bandwidth while its appliances reach more than 90% of LPDDR5X throughput, and that the delivered-bandwidth advantage covers the deficit in peak performance. No peak figure for either memory type appears beside it, so the trade cannot be checked from this reporting. The architecture is described more concretely: a systolic array whose modules keep model weights in co-located memory, separate units for activation functions, and CPU cores Positron calls a "programmable escape hatch" for work the fixed units cannot do. Two published specifications sit oddly together. A Titan holds up to 18.4 terabytes at 23.68 terabits per second, about 2.96 terabytes per second, so one pass over the full pool takes roughly 6.2 seconds. A 32-trillion-parameter model on that box works only if each token reads a small slice of the weights. Tape-out is planned for the end of 2026, so there is no silicon, no toolchain and no benchmark to run against.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence28

0100

One outlet, one source, no measurement

Two copies of a single SiliconANGLE story, an hour apart, are the entire record. Every technical figure in them traces to Positron: the 30% and 90% utilization numbers, the 18.4 terabytes and 23.68 terabits per second, the 32 trillion parameters, the 26 times tokens per dollar. The firmest facts are the ones the company controls anyway, its own raise and its own roadmap, and even those carry no investor voice. What is checkable is the arithmetic on the published specifications, which we did.

Adoption4

0100

Nothing shipping

Titan and Asimov are pre-production by Positron's own account. Tape-out is set for the end of this year, mass production for the second half of 2027, and the only dated event in this reporting is the funding announcement itself. No customer, pilot, cloud partner or delivered rack is named anywhere.

Hype gap

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

invest
Positron's second tranche pays a 16 per cent step-up inside the same announcement
2 publishers
build
Samsung puts MAC trees in every LPDDR5X bank because HBM costs too much
1 publisher
product
Celero raises $275m at a $3bn-plus valuation with its coherent DSP still short of production
1 publisher
build
Neither can build it alone: Korean memory and Taiwanese packaging now set the accelerator clock
1 publisher

Positron bets $875m that smartphone memory can carry inference

What happened

Why it matters

What to watch

Claim ledger

Sources

Topics and entities

Topics

Entities

Claim ledger

Topics and entities

Topics

Entities

Perspective Coverage

Reality

Sources

What happened

Why it matters

What to watch

Claim ledger

Sources

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Related stories

Claim ledger

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Perspective Coverage

Reality

Related stories

Sources