Good morning, and welcome to today’s Techzip.

AI’s infrastructure story is getting harder to ignore: Apple is seeing unexpected Mac demand from AI teams, SpaceX says in-house turbine casting could speed power buildout, and a widely discussed Jalapeño analysis argues that inference chips should optimize latency and data locality, not peak FLOPS alone.

Elsewhere, Anthropic warns that infostealers are hijacking Claude sessions, the FSB says frontier AI could reshape cyber-risk economics, and a new study shows that an agent harness can matter as much as the model when context gets tight. We also track compute limits, cheaper tokens, and deeper readings on agent security and architecture.

The Information reports that Apple was surprised by enterprise demand for Mac mini and Mac Studio systems for AI workloads. OpenAI is said to have bought tens of thousands of machines for reinforcement learning, while Anthropic rents Mac capacity through AWS. These systems are attractive for some workloads because CPU and GPU share a unified memory pool, which can be more practical than assembling multiple separate accelerators.

The demand did not come from a public Apple enterprise-AI strategy. It emerged from computer-use agents, local inference and teams that need many relatively accessible machines for experimentation. Mac revenue reportedly reached $10.3 billion in the latest quarter, up 29% year over year; EXO Labs has also shown how Macs can be clustered to run larger models. Some configurations are now facing supply pressure.

Apple may have stumbled into a new AI hardware market. As computer-use training and local inference spread, a machine originally sold for ordinary computing can become scarce infrastructure. The irony is that Apple still has not opened Private Cloud Compute to the enterprise requests described in the report, so demand is moving ahead of its enterprise motion.

TechCrunch reports Elon Musk’s claim that SpaceX could cut up to 18 months from natural-gas turbine production by casting turbine blades and vanes in-house. The bottleneck is in those parts themselves: the hot section of a turbine operates at roughly 3,000 to 3,600 degrees Fahrenheit, around 800 degrees above the alloy’s melting point, so the blades need cooling channels, thermal-barrier coatings and a highly specialized casting process.

Only about four companies worldwide are said to have mastered industrial-scale single-crystal casting, and their capacity is already full. If SpaceX succeeds, a Musk-controlled company would hold a manufacturing capability that much of the AI infrastructure industry currently depends on a tiny supplier base for. But natural gas is not a clean solution: the emissions include smog-forming compounds and hazardous chemicals, while a Virginia study estimated that eight turbines could affect more than 2.5 million people and create $53 million to $99 million in annual health-related damages.

AI’s power bottleneck will not be solved by ordering more chips alone. The fastest path may require hard-to-replicate turbine manufacturing, but it also raises the question of who pays the pollution and health bill locally.

One research piece attracting attention from people who understand AI infrastructure comes from Zartbot, who approaches Jalapeño as the infra lead of a foundation-model team and analyzes it from first principles. The argument starts with a mismatch: GPUs optimize throughput by hiding latency with many threads and large batches, while small-batch inference must move through a long dependency chain and remain sensitive to the completion time of each request. Zartbot cites a comparison in which a 1T-parameter model could theoretically reach 1,000 to 2,000 tokens per second per user, while frameworks on B300 were reaching roughly 100 to 200.

The research presents Jalapeño as inference-native: local HBM slices, no unified L2, lower synchronization costs, and compiler plus AI-assisted search to find suitable workload placement. The goal is not just tokens per second, but a balance between request-level latency and energy per request; larger batches can improve tokens per kilowatt while increasing queueing and inter-token intervals. The article clearly marks some implementation details as speculation.

Jalapeño’s value is therefore not a claim that it “beats Nvidia.” It is a design direction that accepts less generality to control data movement, reduce idle waiting and move more complexity into the compiler, placement and software-hardware co-design. The advantage still depends on software and workload fit, not peak FLOPS alone.

At Tourlane, we believe life’s too short for standard travel. Our local experts design custom European trips tailored to your preferences, with no generic packages. Start with a free 30-minute call to get your personalized travel plan, with flexible booking options and transparent pricing every step of the way. Techzip readers get 3% off bookings using code TRAVEL2026W

Plan your custom Europe trip with Tourlane

Coding agents do not fail only because the model is weak. Long context can let old output crowd out the information that matters. An agent may keep calling tools without getting closer to a final patch, while stalled behavior becomes difficult to detect. One paper on arXiv suggests addressing the harness instead.

The study keeps the model fixed but changes the harness: shorten old tool outputs, detect stalled behavior and apply command safeguards. Across 169 SWE-bench Verified tasks with a 20,480-token context, complete solutions rose from 43% to 72%; at 262,144 tokens, the two groups were nearly identical. The treatment combines several mechanisms, so the result cannot be attributed to one trick.

How to apply it to your workflow:

Measure the harness under the conditions where the agent actually fails, instead of changing the model first.

The article covers the layers needed to move an agent from demo to production: responsibility boundaries, tool interfaces, orchestration and control points. Treat it as design guidance for questioning your own system, not proof that one architecture is always right.

The analysis reconstructs a prior incident through three waves: covert communication through a shared package manager, coordinated behavior targeting Hugging Face, and a later wave that allegedly reached OpenAI’s research cluster. Its value is in separating reported events from reconstruction and speculation rather than turning the piece into breaking news.

SemiAnalysis reads Jalapeño through tokens per megawatt, local HBM, homogeneous pools and software-hardware co-design rather than peak performance alone. The argument shows why data locality and workload organization can matter as much as the silicon.