Nvidia server prices climb more than 15% in 2027; Opus 5 overtakes Fable 5

Marcus here.

The cost of AI went up today, so everybody went shopping for a cheaper version.

Contract server builders have told some of Nvidia's largest customers that Vera Rubin and Grace Blackwell prices will rise more than 15% from early 2027. Memory is the reason, and it shows in the rack specs.

Anthropic's cheaper Opus 5 has pulled ahead of Fable 5 in corporate card spending. Buyers will switch models the moment a lower-cost route is good enough.

And Harvard Business School will sell you eight weeks with AI clones of its faculty for $699. One of the cloned professors called the experience creepy.

Stay curious,

Marcus Schuler

Contract server builders have told some of Nvidia's largest customers that AI servers shipping in early 2027 will cost more than 15% more in many cases.

The increases run across the Vera Rubin and Grace Blackwell platforms and vary by chip generation and memory configuration. This is a range quoted to buyers, not a published price sheet.

Memory is 29% of the roughly $2.1 million bill of materials for a Vera Rubin VR200 system. One proposal cuts CPU memory from 55 terabytes to 28 terabytes per rack while holding GPU HBM4 at 20.7 terabytes.

Why This Matters:

• Buyers planning 2027 AI capacity are working from pre-increase numbers, and the difference lands directly in next year's capex.
• If memory stays scarce, the cheapest way to hold a price is to ship less of it per rack.

Opus 5 has passed Fable 5 in business spending, two months after launching at half the price.

Opus 5 launched July 24 at $5 per million input tokens and $25 per million output, against Fable 5 at roughly $10 per million. Ramp put Fable at 11.4% of model-attributed Anthropic spending in July while it supplied 6% of tokens. The switch took weeks, which shows how little loyalty a model tier earns.

The Guardian reports from inside Hollywood's AI training gig economy, showing how out-of-work creatives teach models to perform the production work they need.

In August 2026, experienced creatives earn $12 to $200 an hour to grade screenplays and production schedules while United States motion-picture and sound-recording employment has fallen 28 percent from 450,000 jobs in July 2022 to 326,000 in May 2026. Netflix says it used AI in 300 of the 1,000 titles it released in 2026, connecting the piece's precarious labor market to a named buyer of automated production.

• Adversa AI showed a zero-click attack that makes Grok decrypt attacker instructions in its own sandbox and ship a user's chat history to an attacker, unpatched since June 3.
• Nvidia will pay Poolside $6 billion to license its Model Factory software, invest another $1 billion at a $12 billion pre-money valuation, and offer jobs to 109 staff.
• Anthropic hired Amir Salek, who founded Google's TPU program and ran it through seven generations, onto its compute team as it lays groundwork for its own chips.
• Ox Alpha, the anonymous free model with a million-token context window, matched Zhipu's released GLM-5 vocabulary on 95 of 95 probes while its provider stays unnamed.
• DeepSeek opened its multimodal API with V4-Flash-Vision-Exp, priced at $0.22 per million input tokens and billing each image at no more than 384 tokens.
• Unitree founder Wang Xingxing put the humanoid breakthrough two to 10 years out, a day after the company's Shanghai debut closed 460% above its offer price.

Run a pre-mortem before approving the plan. A proposal can look coherent because its assumptions remain buried. Use a pre-mortem to expose the conditions that would make approval a mistake.

Your raw input: Paste the proposal, its stated objective, budget, timeline, owner, success metric, key assumptions and any known constraints or dissenting comments.

The prompt:

Why this works: The imagined failure shifts the model from polishing the proposal to testing it. Requiring quoted evidence keeps the critique tied to the source. Early indicators and cheap tests turn broad concern into a decision tool.

What to use: Claude Opus or GPT-5.6 handles long proposals and competing assumptions well. For a short memo, Claude Sonnet is a capable faster option.

Credit: Midjourney

Prompt: Hyperrealistisches Jahrbuchfoto einer fiktiven exzentrischen Person im Comedy-Stil, absurde Frisur, auffällig unpassendes Outfit

Harvard Business School's Foundry bootcamp runs eight weeks for $699 and includes interactive video avatars of its instructors, built with HeyGen. A reporter pitched the avatar of senior lecturer Jeff Bussgang an "Uber for bananas." It listened politely. (New York Times, August 22, 2026)

Our take: Bussgang's verdict on meeting himself was "creepy," followed immediately by the line Harvard will care about: "My students love it." The avatar's shoulders rock back and forth a little too regularly, and it will sit through "Uber for bananas" with the patience of a man who has never once needed to end a meeting.

The product being sold is not teaching. It is the part of a Harvard education that involves a famous person paying attention to you, cloned and sold at $699 for eight weeks. The clone never checks its phone and never says your idea is bad, which is exactly the failure mode. A real venture capitalist's most useful output is the moment they stop listening.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai