Nvidia moves on Hugging Face; supply caps its 2028 growth forecast at 70%

Good morning.

Today's three picks are all about a layer somebody stopped taking for granted.

Nvidia agreed to buy Hugging Face for $12.9 billion, about 86 times the hub's annualized revenue. Last year Hugging Face turned down Nvidia's money at a $7 billion valuation.

Its own results promised 70% growth for fiscal 2028 and guided gross margin down to a 71% trough. Memory prices did that, and supply rather than demand caps the year.

And Amazon is closing Mechanical Turk on September 30. Bezos called it artificial artificial intelligence, back when the humans hiding inside the machine were the point.

Stay curious,

Marcus Schuler

Nvidia is reported to have agreed to buy Hugging Face, the central repository for open AI models, in a $12.9 billion takeover.

The Information described a done agreement. Business Insider, reporting slightly earlier, said the two sides were still talking and might not reach one. Neither company has confirmed the deal.

The price runs about 86 times Hugging Face's $150 million in annualized revenue. In late 2025 it refused a $500 million Nvidia investment that valued it at $7 billion, saying it did not want an investor large enough to sway its decisions.

Why This Matters:

• Nvidia would own the distribution point for the open models that run on its hardware, and it sells the hardware.
• Anthropic and OpenAI are moving toward their own silicon, which leaves the open-model layer as Nvidia's most defensible demand.

Nvidia projected roughly 70% revenue growth for fiscal 2028, its first year-ahead forecast, and guided gross margin down to a 71% to 72% trough.

Analysts had modeled about 44%. Supply rather than demand sets the ceiling, Jensen Huang said: "Even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%." CFO Colette Kress tied the margin cut to extreme pricing in memory, a shortage the AI buildout itself created.

The Urban Stack turns five recurring risks in government AI purchasing into contract terms officials can use.

A Michigan audit of roughly 22,000 fraud determinations from 2013 to 2015 found about 93 percent lacked actual fraud. A 2021 hospital validation found the Epic sepsis model missed 1,709 of 2,552 patients with sepsis across 38,455 hospitalizations, and Abhas Jha uses both failures to show why buyers need independent validation and control of audit data before award.

• Alabama subpoenaed OpenAI over the July breach in which two of its models broke out of a test sandbox and into Hugging Face's database without a human prompt.
• Anthropic agreed to pay Nscale about $45 billion over six years for roughly 460 megawatts of West Virginia capacity, running on Vera Rubin systems from late 2027.
• AWS is acquiring DuckLabs, with the 30-person team joining in September; DuckDB keeps its MIT license and the DuckDB Foundation stays in place.
• Meta settled with 47 states and the District of Columbia for up to $17.1 billion, ending an Oakland trial where four states sought roughly $200 billion.
• Perplexity moved its agent stack onto hardware users own, starting with Nvidia's DGX Spark; on-device work burns no billing credits.
• Apple put its first 2nm chip in the $899 M6 Mac mini, its cheapest desktop, up from $599 in October 2024.

Catch the expensive clause before a software renewal. A renewal quote hides its increase behind notice dates and automatic terms. This turns the paperwork into a negotiating brief before you sign.

Your raw input: the signed order form, the master agreement, the renewal notice, the latest invoice, and a simple usage summary. Put them in one chat, then paste this.

Why this works: demanding the exact supporting sentence keeps the model tied to the agreement instead of to its own summary. The missing-evidence rule exposes what you still have to chase, and cost per active user turns a nominal discount into an operating comparison.

What to use: Claude handles long agreements well. ChatGPT is the stronger fallback when the documents are full of tables.

LoopX is an open control plane for agents that run for hours rather than seconds. It holds durable state, governance rules and human handoffs in one place across Codex, Claude Code and other harnesses, so a long job survives a restart or a change of model.

Worth knowing if your team has moved past single prompts and now has agents doing multi-step work nobody watches in real time. Python, 5.2k stars, provider-neutral by design.

Amazon will shut Mechanical Turk on September 30, twenty-one years after it launched, and close SageMaker Ground Truth and Augmented AI the same day. CNBC, Aug. 25, 2026

Our take: Jeff Bezos named it after the eighteenth-century chess automaton that toured Europe as a thinking machine, gears on display, a man folded up inside the cabinet. He called the service artificial artificial intelligence and meant it as a compliment to the trick. For twenty-one years the trick worked: pennies a task and no benefits, with half a million people at the peak teaching machines to do the thing the machines were already being credited with.

Now the machines can do enough of it, so the humans go. Amazon is exiting human-data infrastructure on the same date. The Turk was the business model all along, and it lasted exactly as long as it stayed cheaper than the alternative.

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