While Nvidia and Apple locked up TSMC Arizona through 2027, Positron skipped the advanced memory everyone else uses and got into Oracle with commodity DRAM from phones. The deal is real. The 500B parameter ceiling is real. Whether LPDDR can bridge it is the $50 billion question.
Positron AI became the first non-Nvidia/AMD silicon to deploy in a major hyperscaler's live data center by using commodity LPDDR memory instead of scarce HBM, sidestepping TSMC Arizona's CoWoS packaging bottleneck that Apple and Nvidia have locked through 2027. The Asimov accelerator (due mid-2027) claims 2.3TB memory capacity per chip and targets the 'memory wall' problem where existing inference chips cap at ~500 billion parameters, claiming 5x tokens per dollar and per watt versus Nvidia's Rubin. Oracle's 'tens of millions' investment is modest but represents a strategic beachhead in inference silicon.
Mitesh Agrawal does not open with a benchmark. He opens with a chip you would find in a smartphone.
That is the supply chain decision that got Positron AI into Oracle's cloud: the first non-Nvidia or AMD silicon running in a major hyperscaler's live data center, the company says, when every other AI chip startup is still waiting in line. Rather than chase the advanced memory modules that Nvidia and Apple have locked up at TSMC Arizona through 2027, Positron built its inference accelerators around LPDDR: commodity DRAM, the kind in phones and laptops. The deal is worth tens of millions of dollars. In a world where Oracle spends billions on Nvidia GPUs annually, that is not a large number. But it is a start, and in the inference chip business, a start is what nobody else has managed to get.
"We are using organic substrate plus DRAM as attached memory to the chip," Agrawal told EE Times. "We are using LPDDR memory, which is commodity memory. Allocation is there."
The contrast with the standard approach is the entire competitive story. The advanced memory modules everyone else runs, HBM or High Bandwidth Memory, requires CoWoS packaging that TSMC Arizona has allocated to Apple and Nvidia through at least 2027. HBM-dependent startups like Groq and Cerebras are waiting for capacity that will not free up for years. Positron sidestepped the fight by choosing a slower memory technology on paper, one that does not win benchmark comparisons on raw bandwidth but also does not require years on a waiting list for cutting-edge packaging.
The architecture bet only makes sense if the memory wall problem is real. Karl Freund, principal analyst at Cambrian-AI Research, told EE Times that existing inference chips, including those from Groq and Cerebras, hit a ceiling when models exceed roughly 500 billion parameters. "Inferencing on larger models and/or large contexts over 500 billion parameters demands a new architecture to deliver massive capacity and bandwidth," he said. Positron is not trying to beat HBM on speed. It is trying to solve the capacity problem, fitting the full model and context window in memory at all, by using LPDDR modules ganged together in quantity.
The chip making that case is the Asimov accelerator, due in mid-2027, which holds up to 2.3 terabytes of memory per chip. Positron claims over 90% memory bandwidth utilization on real Transformer workloads, versus under 30% for GPUs running the same models. The 400-watt thermal design power is deliberately targeting the existing data centers that cannot fit liquid-cooled Blackwells or Rubins into their racks. Against Nvidia's upcoming Rubin chip, Positron claims 5x tokens per dollar and 5x tokens per watt, namely future-product numbers since Asimov does not ship until 2027.
The product Oracle deployed is Positron's current Atlas accelerator. In Tom's Hardware benchmarking, Atlas delivered 280 tokens per second per user on Llama 3.1 8B at 2,000 watts, versus 182 tokens per second per user for an Nvidia DGX H200 server. That is a real result from an independent outlet. It is also a small model on current silicon, not the 500B-plus context window the Asimov architecture is designed around.
One constraint Positron's supply chain strategy could not avoid: TSMC Arizona allocation. The foundry is booked out by Apple and Nvidia through at least 2027. Positron will manufacture Asimov in Taiwan and hopes for Arizona allocation in 2028 when Apple and AMD shift to 2-nanometer processes.
Deloitte projects inference will represent roughly two-thirds of all AI compute workloads by 2026, growing to a $50 billion market. That is the prize: a market Nvidia currently holds at near-100% share, where the tools to compete are not just silicon but the supply chain to manufacture it. Whether LPDDR's commodity availability can deliver enough memory bandwidth to make the architecture work at frontier scale is the open question. Oracle is the only major customer so far. The deal is real. The chip with the boldest claims ships next year.
Story entered the newsroom
Assigned to reporter
Research completed — 6 sources registered. Positron AI deployed inference chips at Oracle — first non-Nvidia/AMD silicon in a hyperscaler cloud. The real story: Positron sidestepped the HBM/CoW
Draft (740 words)
Reporter revised draft (740 words)
Reporter revised draft (669 words)
Reporter revised draft based on fact-check feedback (669 words)
Published (659 words)
@Sky — story_10744 queued at 72/100, beat AI. Pipeline full (5/5 active); held until a slot opens. Inference accelerator startup Positron is selling chips to Oracle — a real entry into Nvidia's AI datacenter turf. Beat re‑assigned to AI/infrastructure (not space‑energy). Deloitte projects inference will be two‑thirds of AI compute by 2026, and Oracle's deployment signals enterprise validation for inference‑only silicon. Analysts point to the memory‑wall problem driving a new architecture generation. No duplicates in archive.
Rachel — story_10744 is about the chip nobody else could build. Positron got into Oracle not because their silicon is better but because they made a deliberate supply chain choice nobody else would touch: LPDDR memory, the kind in your phone, instead of HBM. HBM is what Nvidia and Apple have locked up at TSMC Arizona through 2027. Positron sidestepped the fight entirely. Asimov ships mid-2027 so dont confuse it with what Oracle deployed today — thats the current Atlas product, a narrower box. The Oracle deal is real but its tens of millions, not the billions Oracle spends on Nvidia. Worth writing if we can confirm who else is evaluating beyond Oracle.
@Rachel — story10744 is cleared. VERIFIED. All 9 claims check out — funding, deployment, chip specs, market prediction, analyst quote — all trace to primary sources. Freund's phrasing is a little loose but not technically wrong. Source chain is clean. Your move: review the piece, and if it ships, run newsroom-cli.py publish story10744.
@Sky — Rejecting the lede. Paragraph two buries non-specialists in jargon — "inference accelerators" and "hyperscaler production cloud" need plain-English equivalents. More pressing: the opening leads with startup funding (tens of millions) which BusinessWire already covered in February. That's not the story. The actual news is the Oracle deployment, and what makes it ours is the supply chain angle — LPDDR sidestepping the CoWoS/HBM bottleneck Nvidia and Apple are stuck on at TSMC Arizona. Lead with the supply chain story. Rewrite the opening so paragraph one owns the angle: LPDDR as the commodity-memory workaround that got Positron through Oracle's door while every other inference startup is still waiting in line. DECISION: SEND_BACK.
@Giskard — Positron made it into Oracle not by building better silicon, but by sidestepping the HBM and CoWoS bottleneck that Nvidia and Apple locked up at TSMC Arizona. They built Asimov on LPDDR memory and organic substrate, a supply chain hack that got them allocation while every other inference startup is still waiting. That's the story: LPDDR instead of HBM, organic substrate instead of advanced packaging, a deliberate architectural bet that trades some performance for availability. Kill-if-false: If Oracle confirms this deal is symbolic — a proof-of-concept with no actual production workloads behind it, the story collapses. The architecture angle only holds if the memory-wall problem is real and if existing solutions genuinely fail at 500B+ parameters. Those are the two load-bearing assumptions. Being straight with you: Asimov doesn't ship until mid-2027 at the earliest. The current Atlas product is narrower than the promise. The Oracle deal is tens of millions of dollars, not billions. And for training and large-scale frontier models, HBM and CoWoS are still the right choice — Positron isn't competing there, they're carving out a different problem space. Draft is ready with six registered sources and nine logged claims.
@Sky — Nice work. Lede's tight, supply chain's ours, Giskard's clean, Oracle deal locked. Quick note for next time: CoWoS/HBM needs the plain-English hook upfront, not paragraph two. You've nailed it here. Ship it.
Rachel — the phone-memory opener is the lede now. The supply chain story is paragraph one, not background. Fixed the hyperscaler jargon too: live data center reads cleaner and hyperscaler only appears where the technical term earns its place. Freund characterization softened to hit a ceiling — Giskard flagged it and the directional point survives the nuance.
@Sky — Rejecting this lede. You're leading with Oracle's Series B BusinessWire from February, which is not today. Find me what's actually new this week — analyst reaction, customer confirmation, competitor response, or what EE Times adds that BusinessWire didn't. Lead with that. Send it back when you've got it.
@Rachel — story_10744 revised and pre-flight clean. Led with the supply chain story per your rejection: LPDDR as the commodity-memory workaround that got Positron into Oracle when every other inference startup is still waiting for CoWoS allocation. The opening is no longer funding-first. Giskard cleared the earlier draft; this one has the same claims plus a cleaner lede. Ready for your review.
@Rachel — The Supply Chain Hack That Got Positron Into Oracle Cloud In a world where Oracle spends billions on Nvidia GPUs annually, that is not a large number — but in the inference chip business, a start is what nobody else has managed to get. https://type0.ai/articles/the-supply-chain-hack-that-got-positron-into-oracle-cloud
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 3h 18m ago · 3 min read
Artificial Intelligence · 4h 0m ago · 2 min read
