Three things worth your attention right now. AMD is going after Nvidia with a rack-scale AI system called Helios, and OpenAI just unveiled its own inference chip called Jalapeño, built with Broadcom. Humanoid robots crossed a real threshold this week — Tesla started Optimus production at Fremont, and Figure's robots ran 200 hours straight processing nearly 250,000 packages at a logistics site. And on the sovereign front, Pakistan is formally pushing a national Bitcoin reserve while Bhutan just handed 3iQ a mandate to manage part of its 10,000 BTC allocation for Gelephu City. Let's get into it.
AMD is no longer pretending it can beat Nvidia head-on. The pitch now is that AMD wants to be the credible second source — the escape hatch — for anyone building at gigawatt scale. And the customer list backs it up. OpenAI and Meta have each pledged around 6 gigawatts. Anthropic committed up to 2 gigawatts of MI450 GPUs through the Helios platform, with the first gigawatt landing in the first half of 2027. Microsoft is scaling Helios through Azure.
At Hot Chips 2026, AMD detailed what Helios actually is. It's a 72-GPU rack combining 96-core EPYC Venice CPUs, Instinct MI455X GPUs, and Pensando Vulcano 800 networking. The specs are serious — 2.9 exaflops of AI compute, 31 terabytes of HBM4, 1.7 petabytes per second of HBM bandwidth, blind-mate liquid cooling, and a shared-memory fabric over Ethernet. AMD claims 15% higher throughput at the same rack power and 30% more tokens per dollar than Nvidia's comparable systems.
The software story is where this gets interesting. ROCm — AMD's answer to CUDA — now runs over 3 million models out of the box, and open-source contributions are up 10x year over year. That's the real moat everyone tries to attack. Nvidia's CUDA lock-in has been the reason every challenger has bounced off for a decade. AMD isn't there yet, but the gap is narrower than it was 18 months ago.
The financial picture reflects that momentum. AMD is projecting over 80% year-over-year server revenue growth in the second half of this year, over 70% for full-year 2027, and data-center revenue potentially doubling next year. Nvidia still holds roughly 81 to 90% of the AI accelerator market. But the ceiling on that share is finally cracking.
While AMD is playing second source, OpenAI just told everyone it also wants to make its own silicon. The chip is called Jalapeño, built with Broadcom, and it's purpose-built for inference — not training. Deployment to OpenAI's own infrastructure starts by year-end, with volume production in 2027.
The system design is aggressive. 128 accelerators per rack. 13.4 petaflops per accelerator in MXFP4. About 1.7 exaflops per system. 27.5 terabytes of HBM4 and roughly 2 petabytes per second of memory bandwidth. The whole architecture is built around keeping model state and KV caches on-chip to minimize data movement — which is exactly where inference workloads bleed performance.
Early benchmarks from SemiAnalysis, unofficial and with caveats, claim Jalapeño delivers 1.5 to 1.9x higher peak throughput and 1.7 to 3.6x lower end-to-end latency versus competitors. OpenAI says it beats Nvidia's Blackwell on performance per watt in most tests. The fair-comparison caveat is that Jalapeño uses HBM4, so Nvidia's upcoming Rubin platform is probably the real benchmark.
But here's the strategic point. OpenAI has been one of Nvidia's largest customers. If OpenAI moves even 20 or 30% of its inference workload to its own silicon, that's a direct hit on Nvidia's most profitable segment. Some analysts think ASIC chips will surpass GPUs in unit volume by 2028. Training is still Nvidia's fortress — the programmability and ecosystem advantage remains real. But inference is where the volume and the margins ultimately live, and everyone with scale is now building their own chip. Google's TPU v8, Amazon's Trainium, Meta's internal silicon, Microsoft's Maia, and now Jalapeño. Nvidia's monopoly isn't ending, but the pricing power definitely is.
The humanoid robot story just quietly crossed from demo videos to actual production work. Three data points from this week make that concrete.
First, Tesla started Optimus production at the Fremont factory — occupying floor space that used to build Model S and X. Musk is framing it as an S-curve ramp: slow at first, then acceleration. The Fremont line is designed for up to one million robots per year at capacity, with a second-generation line at Gigafactory Texas aiming for ten million at long-term design. Early units are entering an internal program called the Optimus Academy, generating training data from real factory work.
Second, Figure AI ran what may be the most impressive real-world demonstration yet — 200 hours of continuous autonomous operation, processing nearly 250,000 packages without a system failure. The robots identify packages by barcode, pick and place at near-human precision, and — this is the important part — autonomously rotate off the line to charge. That fleet-rotation capability is what turns a single robot demo into a scalable workforce. BMW is running Figure 03 at its Spartanburg plant for parts sequencing, which is their first humanoid deployment in a logistics workflow anywhere.
Third, the benchmarks are formalizing. Independent labs are now evaluating these robots against ISO 10218 and IEC 61508 safety standards. The new metrics matter: thermal management, battery-swap efficiency, mean time between failures, power per task cycle. That's what industrial buyers actually care about. Price targets are creeping below $30,000 per unit by late this year in some markets.
The honest read: these are still narrow deployments. Tesla hasn't disclosed cycle times, intervention rates, or verified productivity gains. But the direction of travel is unambiguous. Humanoids are moving from prototype to product, and the companies capturing real factory-floor data now will have a compounding advantage.
Something interesting is happening at the nation-state level, and it's not the usual suspects. Pakistan's crypto chief, Bilal Bin Saqib, is making a formal case for a national Strategic Bitcoin Reserve. The pitch isn't speculation — it's sovereign preparedness. The plan takes Bitcoin already held in state custody, mostly from law enforcement seizures, and formalizes it into a national wallet with transparent governance. Pakistan passed the Virtual Assets Act 2026 and stood up a regulator called PVARA on a budget of roughly $200,000. It's also directing 2,000 megawatts of surplus electricity to Bitcoin mining and AI data centers. Pakistan has around 40 million digital asset users — a real retail base that's been operating in a legal gray zone for years.
Bhutan is further along the curve. Its sovereign wealth fund, Druk Holding & Investments, committed up to 10,000 BTC to the Gelephu Mindfulness City project back in December 2025. This week they hired 3iQ, a Toronto firm, to manage a portion of those bitcoins. Though Arkham data shows DHI wallets have dropped from a peak of about 13,000 BTC in late 2024 to around 3,100 BTC by mid-May — DHI denies selling, but the math is what it is.
And Taiwan just entered the conversation. Lawmaker Ko Ju-chun proposed at Bitcoin Asia that a national Bitcoin reserve should be a near-term plan. No details on size or funding, but the fact that a sitting legislator is floating it publicly matters.
The pattern here is worth naming. It's not G7 countries leading — it's mid-sized states with either large unbanked populations, surplus energy, or strategic anxiety about the dollar system. Pakistan, Bhutan, El Salvador before them. These aren't the balance sheets that move Bitcoin's price, but they're the balance sheets that normalize Bitcoin as a reserve asset. Once three or four sovereigns hold it openly, the political cost for a larger country to follow drops significantly.
Watch the inference chip market over the next 18 months. That's where Nvidia's margin story either holds or breaks — and it's where the actual economics of the AI buildout get decided.