VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog
← Value Add PulseAIEarly-stage chip talks

Anthropic Is Discussing a New Custom AI Chip With Samsung

Anthropic is in early discussions with Samsung to develop a custom AI chip, The Information reported July 2, though the companies have not finalized specifications, use cases, or how powerful the chip would be. The move follows OpenAI's custom inference processor 'Jalapeño,' announced the prior week through a partnership with Broadcom, and comes as Anthropic separately explores ways to address chip shortages it flagged as early as April 2026.

July 2, 2026 (The Information)
Report Date
OpenAI's 'Jalapeño' (with Broadcom), announced prior week
Rival Custom Chip
April 2026
Anthropic's Chip Concerns Raised
Google, Amazon, Nvidia
Current Compute Partners
Early discussions; specs, use case, power not finalized
Deal Status
TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
July 2, 2026
3 min read
ShareXLinkedInEmail
KEY TAKEAWAYS FOR VCs & FOUNDERS
1

Every major AI lab is now pursuing custom silicon — Anthropic joining OpenAI (Broadcom's 'Jalapeño'), Google (TPUs) and Amazon (Trainium) confirms bespoke chips are now table stakes, not a differentiator

2

Anthropic explicitly maintains that 'a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal' — this supplements rather than replaces existing compute relationships

3

Samsung already manufactures chips for Nvidia and collaborates with Google on chip development, making it a credible foundry partner rather than a speculative newcomer to AI silicon

4

Coming right after OpenAI's custom inference chip announcement frames this explicitly as a competitive response, not an independent strategic decision

TC
The VC Read · Trace's TakeTrace Cohen

Every frontier lab racing to announce a custom chip within days of each other (OpenAI's Jalapeño last week, now Anthropic exploring Samsung) tells you custom silicon has become a mandatory checkbox for credibility, not just a cost play — no lab wants to be the one still standing there as a pure Nvidia customer while its two biggest rivals talk about proprietary hardware roadmaps. The fact that Anthropic can't yet say what the chip is for, how powerful it'll be, or how it fits into a rack is the real story here — this reads much more like a defensive press-cycle move than a mature program. For founders building infrastructure or tooling in the AI stack, the durable takeaway is that chip-agnostic software — the kind Together AI and Modular are selling — gets more valuable, not less, every time another proprietary chip enters the mix. Watch whether Anthropic actually names a chip and a use case the way OpenAI did with Jalapeño, or whether this quietly fades as an exploratory conversation.

⚔️ AI Chip Wars →💾 AI Chip Startups →

Anthropic is in discussions with Samsung to develop a custom AI chip, according to reporting from The Information published July 2, 2026, though the two companies have not yet finalized key specifications, the chip's intended use case, or how powerful it would ultimately be. When contacted for comment, Anthropic declined to elaborate beyond stating that "a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy" — language explicitly framing any Samsung chip as an addition to, not a replacement for, its existing compute relationships.

The timing points to two overlapping motivations. First, Anthropic has been exploring ways to address chip shortages since at least April 2026, a persistent bottleneck as frontier labs compete for a limited supply of leading-edge AI silicon. Second, and more immediately, the move follows OpenAI's announcement the prior week of its own custom inference processor, code-named "Jalapeño," developed in partnership with Broadcom — putting pressure on Anthropic to demonstrate it has a comparable hardware roadmap rather than remaining a pure customer of merchant silicon from Nvidia and cloud-provider TPUs.

The broader pattern is now unmistakable: every major AI lab is pursuing some form of custom silicon. Amazon and Google both already offer custom AI accelerators (Trainium and TPUs, respectively) to reduce dependence on Nvidia and optimize hardware for their specific workloads. Samsung itself is not a newcomer to this ecosystem — the company already manufactures chips for Nvidia and collaborates with Google on chip development, making it a credible foundry and design partner rather than a speculative outsider trying to break into AI silicon for the first time.

“The broader pattern is now unmistakable: every major AI lab is pursuing some form of custom silicon.”

What remains genuinely unresolved, per the reporting, is basic scope: whether the chip is intended primarily for inference or training, how it would physically integrate into server racks alongside Anthropic's existing Nvidia and cloud-provider hardware, and what performance target Anthropic and Samsung are even designing toward. That level of ambiguity this early suggests the relationship is closer to an exploratory partnership than a committed, specced-out silicon program — a meaningfully earlier stage than OpenAI's already-named Jalapeño chip.

The strategic logic for Anthropic is straightforward even without finalized specs: custom silicon, even for a narrow slice of workloads, gives a frontier lab negotiating leverage against Nvidia's pricing and margin capture, some insulation from GPU supply shortages, and — increasingly — a talking point for investors ahead of Anthropic's own reported IPO preparations, where a diversified, partially self-controlled hardware roadmap reads as a more resilient story than total dependence on merchant silicon.

For founders and infrastructure investors, this is further confirmation that the AI chip landscape is fragmenting into a multi-vendor world (Nvidia, custom hyperscaler silicon, and now custom frontier-lab silicon) rather than consolidating around a single dominant architecture — a dynamic that keeps chip-agnostic software layers like Together AI's neocloud stack, or GV-backed Modular's disaggregated-inference approach, strategically relevant regardless of which lab's custom chip ships first. For LPs watching semiconductor and AI-infrastructure exposure, Samsung landing a design relationship with two of the three leading US frontier labs (alongside its existing Nvidia manufacturing and Google collaboration) reinforces its position as a foundry beneficiary of the AI buildout independent of which specific lab wins the model race.

What to watch: whether Anthropic and Samsung formalize specific chip specifications and a target use case in the coming months, how the chip compares to OpenAI's Jalapeño once both are further along, and whether Anthropic discloses any custom-silicon detail in its own IPO filings given how directly hardware strategy now factors into how public investors will value frontier labs.

ShareXLinkedInEmail
More onAnthropic →OpenAI →

Originally reported by The Information. Analysis and editorial commentary by Value Add Pulse.

← Back to Pulse

Read Next

AI5% equity proposed

OpenAI Proposed Donating 5% of Its Equity to a US Sovereign Wealth Fund

OpenAI CEO Sam Altman proposed donating 5% of the company's equity to a US sovereign wealth fund, the Financial Times reported July 2, with other AI companies expected to make similar contributions so the public can share directly in AI-driven financial gains. The proposal follows a April 2026 OpenAI policy paper arguing such a fund could 'invest directly in AI labs and companies deploying their technology,' and lands alongside competing legislation from Senator Bernie Sanders proposing a 50% one-time tax on systemically important AI companies instead.

AI$145B 2026 AI spend

Zuckerberg Tells Staff AI Agents Haven't Progressed as Quickly as He'd Hoped

Meta CEO Mark Zuckerberg told employees at an internal town hall this week that AI agent development has not 'accelerated in the way' executives previously anticipated, and that benefits from Meta's AI-focused reorganization 'haven't come to fruition yet.' The admission follows a restructuring earlier in 2026 that laid off roughly 8,000 employees (10% of the workforce) while reassigning 7,000 more into AI groups including 'Agent Transformation,' as Meta spends an expected $145 billion this year on AI infrastructure.

AI>99% token reduction

Alibaba's SkillWeaver Framework Cuts AI Agent Token Consumption by Over 99%

Alibaba researchers published SkillWeaver, a framework that decomposes complex agent tasks and retrieves only the relevant tools from a library rather than loading an entire tool catalog into context, cutting token consumption by more than 99% compared to naive approaches, according to VentureBeat reporting July 2. Tested on a custom 300-query benchmark against a library of 2,209 real-world MCP tools, SkillWeaver's feedback-loop technique boosted task-decomposition accuracy from 51% to 92% with a larger model, while reducing per-query context use from an estimated 884,000 tokens to roughly 1,160.

@Trace_Cohen·t@nyvp.com