The body stops at 8000 characters, mid-sentence, cut by the RSS leg
— with no marker of its own. It is clean, well-formed prose that reads as a complete
article. Without completeness, a vector store would answer from this preview with the
same confidence as from the full piece. {"complete": false, "kind": "hard-cut", "marker": "8000-char fixed-width cut"}
24
frontmatter fields
13
fields with an actor
8,000
body characters
1
narratives
34
entities
0
tickers
Trust — derived from the actor, never stored
H
Human attestedstrongest — a person signed this · 0 field(s)
M
Machine deterministicfilterable as fact · 10 field(s)
D
Model deriveda reading, not a fact · 2 field(s)
U
Unverifiednobody attested — unknown, not none · 12 field(s)
OKF §7 gives three actor forms —
human:<id>, process:<id>, <producer>/<version>
— and the consumer derives the tier from the form. No parcel stores a tier: storing one would be a
second, conflicting answer to a question the standard already settles.
tickers: [] carries no actor, so it reads as unknown, not none.
Can AMD break the CUDA Moat? AMD Advancing AI 2026
M
process:okf-parcel-v2.map
author
Dylan Patel
M
process:okf-parcel-v2.map
published_at
2026-07-25T00:33:23Z
U
—
retrieved_at
2026-07-25T00:33:23Z
M
process:okf-parcel-v2.map
status
stable
U
—
media
[]
M
process:okf-parcel-v2.map
Enrichment — written by S2
field
value
tier
actor
summary
SemiAnalysis upgrades its standing view on whether AMD can close the CUDA gap: from a 0% chance in its first AMD software article, to a non-zero chance in AMD 2.0, to a great chance of success now, conditional on two unr …
Narratives — machine-deterministic, filterable as fact
ai-memory-storage
A narrative names a subject, never a position size. Fund allocations are out of scope here by operator ruling and by construction: the theme taxonomy is constant, the book is not. Enforced by a negative control — --check scrambles every NAV weight in the theme table and asserts no label moves, and that control is itself proven to fail when the dependency returns.
SemiAnalysis upgrades its standing view on whether AMD can close the CUDA gap: from a 0% chance in its first AMD software article, to a non-zero chance in AMD 2.0, to a great chance of success now, conditional on two unresolved risks. Risk one is the Helios rack ramp: without a cableless tray design (which Nvidia's Rubin Oberon rack is adopting) and with a weak SerDes forcing up to 85% of the backplane to be retimed via more than 550 Broadcom ethernet retimers per rack, production is ramping slowly against backplane reliability problems. Risk two is internal capacity: AMD engineers report a persistent lack of stable GPU clusters for software development and for automated-testing CI, which caps software quality and blocks AMD from exploiting AI coding agents, since each agent itself needs GPUs and a testing loop. Commercial momentum runs the other way: Anthropic has publicly announced it will deploy 2GW of AMD chips, Microsoft has reversed the post-MI300X exit that made it skip MI325X and MI355X and will now deploy MI455X Helios (with OpenAI as the expected end customer on Azure), and AMD has a Cerebras deal for prefill-decode disaggregation for fast interactive inference. A stock-option structure gives Meta and OpenAI close to a 105% equity rebate, fully triggered when AMD stock reaches $600 and enough compute is bought, which the author argues drives Helios cost per million tokens to practically negative. Specific gaps named: Kubernetes Inferencing Pollara NIC CI sits at 0% parity with Nvidia's ConnectX nightly CI, and vLLM gating regressed when leadership pulled clusters from the internal vLLM team, missing the target of 90% CUDA parity by Advancing AI 2026. The piece also notes the MI455X (gfx1250) instruction set is a clone of Hopper SM90's, and that much of the ROCm stack, including the MoRI collective and UMBP KVCache offloading work, is built in Shanghai. The retained body is cut mid-sentence and does not include the article's Helios architecture, networking, software-stack or TCO sections.
Body — publisher chrome stripped
When we published our first AMD software article, we gave AMD a 0% chance of closing the gap with Nvidia in AI accelerators. Software was broken, progress was unexciting, and we were the top bug submitter for many months with dozens of AMD engineers triaging our bug reports.
Six months later, in our AMD 2.0 article, we took the non-consensus position of upgrading from 0% chance to a much more meaningful chance at success. We published this opinion when sentiment towards AMD was sitting at rock bottom, and when most of the market thought we were were being far too optimistic towards AMD.
That view was based on our observations that AMD has the leadership that can create change rather than suffer from committee style leadership. Lisa quickly hopped on a calls with us and has since then implemented many of our suggestions. We saw an AMD that had finally recognized the importance of software and had a sense of urgency, moving in the right direction even if the destination was still far off. Since then, the signal has only gotten stronger.
Based on our experience on AMD software stack this year, we update our view again from non-zero percentage chance to now a great chance of success as long as AMD solves the two major risks we outline below. It is important to highlight that just because AMD gains market share, that doesn’t mean that Nvidia will do poorly. The pie is growing rapidly for everyone, and Nvidia will continue to massively grow revenue. AMD poses potential competiti
▼ cut here at exactly 8,000 characters, mid-sentence. The article's Helios architecture, networking, software-stack and TCO sections are not in this parcel.