posts · 2026-07-07

This is straight from Hugging Face's official blog — both sources are republishing the same announcement, no independent testing or third-party angles yet. The change: previously, if you used SkyPilot to run training on AWS with data in S3, then spun up GPU instances on GCP, you'd pay egress fees pulling that data across clouds. Now, store it on Hugging Face Storage, SkyPilot mounts it directly, and cross-cloud moves skip the egress charge. For teams that hop between clouds chasing cheap spot GPUs, this is a real cost win. But don't read it as "multi-cloud storage solved" — Hugging Face Storage still sits on AWS underneath; they've just set up zero-egress peering. I haven't seen pricing comparisons, latency numbers, or any sign other cloud providers will offer similar setups. What I'd actually watch for: someone posting real cross-cloud training throughput benchmarks. That's the test.

Tomasz Tunguz ran the numbers: over the past 12 months, OpenAI, Microsoft, Anthropic, Amazon, and Google Cloud collectively committed $9.75 billion to forward-deployed engineering—putting their own engineers inside customer orgs to get AI actually working. That's one quarter of Accenture's annual labor cost. This isn't experimental anymore. Three models are emerging. Microsoft and Amazon fund FDE from existing headcount—no outside capital, fast and controlled. OpenAI and Anthropic created standalone entities backed by PE firms like TPG and Blackstone; OpenAI also acquired Tomoro, a 150-person consultancy. Google Cloud took the third path, committing $750 million to a partner fund instead of building direct. Tunguz's core argument: FDE is becoming a moat. Once a customer team is trained on one lab's stack, the organizational friction of switching is brutal. Embedded engineers also see proprietary workflows, data schemas, and failure modes that no API call reveals—intelligence that feeds directly back into model tuning. I'd discount the moat claim a bit. Embedding engineers definitely creates stickiness, but if a competitor's models pull far enough ahead, customers will still switch. This looks more like a land-grab lock-in fee than a permanent barrier.

OpenRouter ran five models on MMMU-Pro Vision and the takeaway is blunt: low-res images on reasoning models cost more, not less. gpt-5.5 hit 79% on auto, dropped to 65.2% on low, and the per-question cost went up from 4.5¢ to 5.1¢. The model burned 1.6× more reasoning tokens trying to decipher blurry inputs, and those extra output tokens ate the input savings. Non-reasoning models gpt-5.4-mini and gpt-4.1 did save money on low, but lost 9.7 and 17.4 accuracy points respectively. Charts and graphs benefit most from auto: gemini-3.1-pro jumped from 78.6% to 91.7%. The practical advice is solid—send clear images and dial down reasoning effort if you want to cut costs. Don't starve the model of pixels.

This piece is worth reading because it names the most common mistake in AI cost-cutting: reaching for a cheaper model before understanding what the tokens are actually buying. The split is clean. Internal efficiency costs (Copilot, Cursor, internal assistants) should be attacked by killing idle seats and capping agent loops first—unused licenses and infinite retries burn more than model pricing ever will. Customer delivery costs (support AI, Duolingo Max video calls) need cost-per-outcome tracking; swapping in a cheaper model that drives up handoff rates destroys gross margin faster than token savings help. Two data points ground this. CMB burns 33B tokens daily, yet AI coding takes only ~5% of compute—a reminder that enterprise token volume often lives in support and ops, not the dev tools that dominate tech discourse. Doubao hit 180T daily calls, but analysts question how much is paid production vs. free trial quota. Same skepticism applies to internal cost reviews: if you can't separate production from pilot from test traffic, you're cutting blind. I'd use this as a governance checklist. Attribute costs to teams and outcomes first, negotiate model pricing last. Reverse that order and you'll probably cut the wrong thing.

This one's worth opening because it moves diffusion-based text generation from a research curiosity to something that looks engineerable. Nemotron-Labs-Diffusion runs three modes in one model: standard token-by-token AR, one-shot denoising for whole segments, and a "self-speculation" mode where diffusion drafts and AR verifies. The training objective is joint—diffusion handles lookahead planning, AR supplies left-to-right priors. The headline number: the 8B model decodes 6× more tokens per forward pass than Qwen3-8B at comparable accuracy, yielding 4× higher throughput on a GB200 with SGLang. I'd discount this on two fronts. First, the paper doesn't mention open-weight plans or a release timeline—it's a preprint. Second, latency and memory behavior on non-GB200 hardware isn't detailed, so don't assume your A100 cluster will see 4× gains. But the direction is real: if diffusion-drafting plus AR-verification stabilizes, high-concurrency inference costs could drop meaningfully.