Compute Economics

Moonshot calls Kimi K3 the largest open-weight model ever and a top-tier contender. Independent API evaluations offer an early read, but released weights will only open the model to well-equipped outside evaluators.

SK Hynix's CEO told Reuters that 2027 will be the worst year yet for memory supply, with customer demand forecast to exceed its own capacity beyond 2030, on the day its ADRs began trading on Nasdaq. TrendForce has reported that SK Hynix removed price caps from long-term supply contracts, leaving customers more exposed to a prolonged rise in memory prices.

Jeff Bezos's Prometheus closed a $12 billion round at a $41 billion valuation with no product and no benchmark. The dollar figure is not the story. The language model labs were handed a free, infinite text corpus; the physical economy generates almost no scrapeable data. Read that way, the roughly $18 billion raised so far is the cost of manufacturing a proprietary physical dataset that does not otherwise exist, which is also why the checks came from banks and asset managers, not just venture funds.

Amazon signaled it may sell Trainium chips to outside data centers, following Google's TPUs into the merchant market. Anthropic and Meta are each moving toward Samsung for custom silicon of their own, one in early talks and one reportedly negotiating a deal, evidence that Nvidia's moat was always software, not chips.

With the unveiling of Jalapeño, its first custom AI accelerator, OpenAI has completed a transformation few predicted when it launched ChatGPT three years ago. The company now designs its own chips, builds its own data centers, and is developing its own consumer hardware. The question is what happens to everyone who supplied those things before.

SharonAI Holdings reported $294,014 in quarterly revenue before signing a six-year, $4.88 billion compute deal with NVIDIA - made possible because NVIDIA itself provided credit support. The arrangement is a small, legible version of a pattern NVIDIA has been building at far larger scale across OpenAI, CoreWeave, Nebius, and others, raising pointed questions about what counts as independent demand.

The contracts being signed this quarter will determine who can afford AI infrastructure in 2028. Dell's record Q1 FY2027 earnings reveal how memory scarcity is quietly splitting the AI buildout into two tiers: enterprises with procurement leverage to lock in multi-year OEM agreements, and everyone else.

The industry just commissioned a study to prove data centers don't raise your power bill. That it needed to is the story. In the first quarter of 2026 alone, local opposition killed at least 20 projects and $41.7 billion in planned investment, even as Morgan Stanley warns the grid is already heading 44 gigawatts short. The binding constraint on the AI buildout has moved from the balance sheet to the county planning board.

The wetware computing industry is betting billions that living neurons can outperform silicon. A new organism called the neurobot, which grew its own nervous system from scratch with no evolutionary history and no instruction, may be the most radical proof of concept yet, and it raises questions that AI researchers cannot ignore.

Samsung Electronics crossed a $1 trillion market capitalization on May 6, posting a 15% single-session surge after Q1 2026 operating profit of 57.2 trillion won - up 756% year-over-year. The numbers signal something more fundamental than a cyclical peak: AI has structurally transformed memory from a commodity into a constrained, strategic resource.

Nvidia B300 servers now sell for around $1 million in China - nearly double the U.S. list price. The price surge is a direct consequence of two converging pressures: the H20 export licensing requirement that cost Nvidia $4.5 billion, and a federal indictment that dismantled the grey-market supply chain that had kept restricted hardware flowing to Chinese buyers.

Meta beat earnings expectations and delivered its fastest revenue growth since 2021. Then it raised its 2026 capex forecast to $145 billion and watched its stock fall. The company's problem isn't the numbers; it's that it still can't answer the most basic investor question about them.

Alphabet, Microsoft, Amazon, and Meta reported Q1 2026 results on April 29 that collectively delivered the clearest evidence yet that AI infrastructure spending is generating real cloud revenue. The outlier was Meta, whose strong earnings were overshadowed by a capex guidance range raised for the second time this year, with no concrete product milestone attached to the ceiling.

Intel posted its strongest quarter in years, with revenue beating Wall Street by $1.3 billion and its data center and AI unit up 22% year over year. The real story is structural: the AI infrastructure buildout is quietly rehabilitating the CPU, and Intel finds itself holding assets no one expected to matter this much.

On April 8, Elon Musk listed seven models in simultaneous training on Colossus 2 and captioned the post "Some catching up to do." The cluster burns 400 megawatts, runs on an estimated 550,000 NVIDIA Blackwell GPUs, and is training a 10-trillion-parameter model. The question is whether scale alone can close the gap.

Google Research has published TurboQuant, an algorithm that cuts the memory cost of running large AI models by at least sixfold - with no accuracy penalty and no retraining required. Memory chip stocks sold off sharply. The sell-off misread what the research actually says.

Cerebras and AWS are deploying CS-3 wafer-scale systems inside Amazon data centers, pairing them with Trainium in a disaggregated inference architecture available through Amazon Bedrock. The setup targets the memory-bandwidth bottleneck that limits GPU-based decode, promising thousands of output tokens per second for agentic workloads.

NVIDIA's GTC 2026 keynote unveiled a trillion-dollar order outlook, the Vera Rubin platform, Dynamo 1.0 as an inference operating system, and a landmark Meta partnership; together they make the case that the future of agentic AI runs on a single, vertically integrated stack.

At GTC 2026, NVIDIA unveiled NemoClaw, a secure software stack that installs Nemotron models and the new OpenShell runtime onto OpenClaw agents in a single command. The move signals something larger than a product launch: NVIDIA is positioning itself as the indispensable infrastructure layer for the agentic AI era.

Jensen Huang's GTC 2026 keynote crystallizes an ambition that has been building for years: NVIDIA wants to own the entire AI infrastructure stack, from silicon to software to agents. Three headline announcements - the Rubin GPU architecture, a Groq-derived inference system, and the NemoClaw enterprise agent platform - make the case in full.

The Trump administration is drafting rules that would require a U.S. government license for virtually every overseas sale of advanced AI chips, regardless of the buyer's location. The tiered framework - covering deployments from under 1,000 chips to installations of 200,000 or more - marks a fundamental break from the Biden era's ally-exemption model, and raises questions about whether chip access is becoming a trade lever as much as a security tool.

NVIDIA's Vera Rubin platform, announced at CES 2026 and entering production this year, promises 10x lower inference token costs and 5x per-GPU compute over Blackwell. This is not an incremental upgrade. It will fundamentally reshape who can afford to build frontier AI.