The Arena

Nvidia tells top customers to expect 15%+ price jumps from early 2027

Editorial

Editorial

— 2 min read
Share
Nvidia tells top customers to expect 15%+ price jumps from early 2027

Two stories today point in opposite directions: the hardware bill for AI is about to get bigger, while small models keep chipping away at the idea that bigger automatically means better.

Some of Nvidia's biggest customers have been told to expect server prices to rise more than 15 percent starting in early 2027, Bloomberg reported on Saturday. The increases cover machines built around the upcoming Vera Rubin platform as well as current Grace Blackwell systems, according to people familiar with the matter, making this the first across-the-board price hike of the generative-AI era rather than a quiet SKU adjustment. It lands days after Samsung said it would raise chip prices by as much as 15 percent on surging AI demand, and hours after Amazon pointed to soaring memory costs as the reason Echo, Kindle and Fire TV got more expensive. The whole buildout has run on one comfortable assumption — each hardware generation makes tokens cheaper to produce — and that assumption is now visibly bending, because memory is the new bottleneck: HBM and DRAM capacity is being pulled toward data centers, and everyone downstream pays. If the companies that buy GPUs by the rack are getting a 15-percent head start on budgeting, expect those increases to show up in cloud pricing and, eventually, in your API bill. We flagged the memory squeeze this week in The AI boom's appetite for memory is outgrowing the industry that makes it — this is what it looks like when it reaches the top of the food chain.

A London lab founded by DeepMind alumni claims its agent beat Anthropic and OpenAI's best at replicating scientific research — running on a model one-tenth the size. Inherent says Faraday, an AI "teammate" built on Alibaba's open Qwen 3.6 (27 billion parameters), outperformed Claude Opus 4.8 and GPT-5.5 at independently reproducing the findings of published papers without being told the answers in advance. Chief scientist Edward Hughes told TechCrunch the interesting part was not the win itself but how they got there: instead of training primarily on descriptions of how science works, the team leaned on reinforcement learning that rewards well-designed experiments — an attempt to teach what researchers call "research taste." Treat the result carefully: it comes from the company's own evaluation, not an independent benchmark. But it rhymes with a pattern we've tracked all month in Prime Intellect: open models nearly match Opus 5 at AI research — small, narrowly trained systems are starting to beat frontier giants on tasks that actually require thinking, not just recall.

What to watch: whether Nvidia confirms the increases as Vera Rubin ramps into production next year — and whether Faraday's replication scores survive contact with independent evaluators.

If a 27B-parameter agent can genuinely do a PhD student's first job, what exactly are you still paying frontier-API prices for? Tell us in the comments.

Read more

Houmo moves compute inside memory — and A20 enters production

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model