Intel's Arc Pro B70 Quietly Beats NVIDIA's Flagship in AI Inference — At a Quarter of the Price
You don't see this every day. Intel's Arc Pro B70 — a workstation GPU nobody really talked about — is reportedly beating NVIDIA's RTX 5090D in DeepSeek R1 inference, pushing over 2000 tokens per second. And it costs roughly a quarter of what you'd pay for Team Green's flagship.
Let that sink in for a second.
I've been running local LLMs on various hardware for the past year — NVIDIA cards, Apple Silicon, even tried CPU inference just for fun. The NVIDIA tax has always felt justified because, well, CUDA is everywhere and the raw compute is undeniable. But this Intel result flips a few assumptions on their head. The B70 isn't even a consumer gaming card — it's an Arc Pro variant meant for workstations, and here it is outperforming the 5090D on an actual open-source model workload.
To be fair, synthetic benchmarks and real-world throughput are two different animals. But 2000+ tokens/s on DeepSeek R1 is nothing to sneeze at. If you're running self-hosted inference at any kind of scale, the cost difference alone makes this worth a hard look. Imagine deploying ten B70s for what you'd pay for two 5090Ds — the math writes itself.
The caveat? Software stack maturity. Intel's oneAPI and OpenVINO have come a long way, but they're still not plug-and-play the way CUDA is. If you're the type who enjoys tinkering with docker images and PyTorch compile flags, this is your playground. If you just want things to work out of the box, NVIDIA still holds that crown.
The Infrastructure Conversation Is Shifting — CPUs Are Back in the Picture
Arm's CEO made a comment this week that stuck with me: AI agents are going to drive CPU demand, not just GPU demand. For the past two years, every AI infrastructure discussion has been "how many H100s do you need?" But as agents get smaller, more specialized, and more distributed, the bottleneck shifts.
I've been running a small fleet of AI agents — one handles my email triage, another watches my GitHub PRs, a third manages calendar scheduling. None of them need an H100. They run fine on a Mac Mini with an M-series chip, humming along at single-digit watts. The Arm CEO's point is exactly this: when you have hundreds or thousands of lightweight agents running across an organization, the aggregate CPU demand dwarfs what a few GPU nodes consume.
Apple's silicon leadership made a similar case — a $799 Mac Mini could honestly be all you need to run your personal agent stack. I've been saying this for months in my local AI setup: not every inference needs a datacenter GPU. My Mac Studio handles a surprising amount of daily agent work without breaking a sweat.
The catch? Apple's memory bandwidth tax. Once you push past a certain number of concurrent agent threads, the unified memory starts showing its limits. For personal use, it's fantastic. For anything approaching production scale, you'll still want dedicated hardware.
LLM Leaderboards Are Losing Their Grip on Enterprise Decisions
Databricks dropped a benchmark study this week that basically confirmed what many of us have suspected: enterprise AI buyers are moving past public leaderboards. The decision criteria have shifted from "which model scores highest on MMLU?" to "which model performs best on my specific data at a cost I can stomach?"
I've watched this play out firsthand. A friend runs AI ops at a mid-size SaaS company, and he told me they evaluated seven different models for their customer support pipeline. The one that won wasn't the top scorer on any leaderboard — it was the one that hallucinated least on their specific product documentation and cost 60% less per million tokens than the flashy alternative.
Real-world performance, cost predictability, and ease of fine-tuning are now the trifecta that matters. Leaderboards are a starting point, not a decision tool.
The Dark Side: First Fully Agentic Ransomware Has Been Documented
Researchers at Sysdig uncovered something unsettling — what they're calling the first documented case of fully agentic ransomware, codenamed JADEPUFFER. The attack is driven entirely by an LLM, targeting cloud servers, and it's smart enough to adapt its behavior mid-operation.
This isn't your typical ransomware that encrypts files and demands Bitcoin. This one uses an LLM to scan the environment, identify the most valuable targets, craft context-aware phishing follow-ups, and even negotiate with victims. It's basically a malicious AI agent with a mission.
The implications are uncomfortable. We've been talking about AI safety in the abstract for years — alignment, bias, jailbreaks. JADEPUFFER makes it concrete: a weaponized LLM that doesn't just generate text but executes a full attack chain. Security teams need to rethink their threat models. Traditional signature-based detection won't catch something that rewrites its own approach every time it runs.
Small Models, Big Impact
Two smaller stories worth noting: Google's Gemma 4 E4B is proving that you don't need a massive model for everyday productivity tasks. It's small enough to run on a laptop, yet handles typical LLM workloads surprisingly well. I tested it briefly on a MacBook Air — response times were snappy, and for tasks like summarization and basic Q&A, it held its own against much larger models.
And on the self-hosted front, one developer shared their experience replacing Gemma 4 with a two-year-old coding LLM for non-coding work — and it actually worked better. The lesson? Sometimes a specialized, older model trained on structured inputs outperforms a newer generalist. Not every problem needs the latest release.
That's the landscape today. GPUs are getting real competition from unexpected places, agents are reshaping how we think about infrastructure, and the security world just got a sobering reminder that AI cuts both ways. The next few months are going to be interesting — keep your local models handy and your security patches up to date.
Thinking about building your own agent stack? PayCalc has some interesting comparison tools for hardware sizing.
Top comments (0)