Memory Prices Jump as Micron Nears Record High, Yanyu Sets AI App Funding Record, and Enterprise AI ROI Remains Unproven

This week, the bottom layer and the top layer of AI infrastructure are moving in opposite directions. At the bottom layer, memory and compute have entered a new round of structural price increases. After server DRAM contract prices rose by approximately 90–95% quarter-on-quarter in the first quarter, they are expected to rise by more than 60% again in the second quarter. HBM capacity for the full year is largely sold out. Micron and SanDisk shares have reached record highs, with year-to-date gains measured in hundreds of percentage points. At the top layer, however, enterprise AI applications and implementation are still having to justify their returns. MIT research shows that 95% of generative AI pilots have produced no measurable profit-and-loss impact. In BCG’s survey of 1,800 executives, only 26% of companies reported substantial financial value from AI investments. These two developments are happening at the same time, placing pressure on both sides of enterprise AI budgets. Unit token prices have fallen by more than one hundredfold over the past two years, yet enterprise AI bills have increased by more than threefold. The main reason is that agentic workflows — workflows that autonomously call tools and execute multi-step tasks — have pushed token consumption per task up by 5 to 30 times. Capital deployment into the application layer has not cooled. In China, Yanyu Technology has just raised nearly USD 300 million, setting a new domestic record for a single funding round in the AI application layer. Moonshot AI’s valuation has increased by roughly sixfold in half a year, reaching approximately USD 30 billion. Companies that have secured funding are increasingly diverging into two categories: those with real cash flow and those still burning capital. What they all face, however, is the same upward-sloping compute cost curve. For enterprise decision-makers, this week’s signal is concentrated in one place: in an environment where model and compute prices are moving in both directions and returns are repeatedly being scrutinized, a control layer that can allocate every dollar of inference budget to productive tasks and switch suppliers at any time is moving from an optional capability to a prerequisite for financial discipline.

Key Events

Memory Enters a Structural Price Increase Cycle: Contract Prices, Equity Markets, and the Upcoming Micron Earnings Release

The first point to clarify is why memory has become the bottleneck in this compute cycle. High-end GPUs used for training and inference must be paired with HBM, or high-bandwidth memory, which is stacked directly beside the GPU and supplies data at high speed. The three major manufacturers — Samsung, SK Hynix, and Micron — have concentrated production capacity toward higher-margin HBM, compressing the supply of conventional DRAM. This has pushed up prices across the spectrum, from HBM to ordinary server memory.

In terms of magnitude, server DRAM contract prices already rose by approximately 90–95% quarter-on-quarter in the first quarter. TrendForce and downstream channels expect a further increase of approximately 63% in the second quarter, with the trend continuing into the third and fourth quarters. HBM capacity for 2026 is largely sold out under multi-year long-term agreements, while SK Hynix has said that its HBM, DRAM, and NAND capacity for 2026 is essentially fully sold. IDC has described this shift as a permanent reallocation of capacity toward AI.

Capital markets have also responded. Micron’s share price approached a record high of USD 1,089 this week, with gains over the past year ranging from 700% to 900% depending on the measurement basis, and a market capitalization of approximately USD 1.2 trillion. SanDisk has risen by approximately 558% year to date, making it the best-performing stock in the S&P 500 this year.

The next key milestone is Micron’s earnings release after market close on June 24. The company previously guided for quarterly revenue of approximately USD 33.5 billion and a non-GAAP gross margin of approximately 81%. The market views this earnings report as a test of whether HBM demand is real and durable, or approaching a peak. Some analysts have also warned of the inherent boom-and-bust cycle risk in the memory sector.

Memory price increases will pass through layer by layer into enterprise bills via server costs, cloud instance pricing, and model inference prices. This cost increase often occurs at the same time as price cuts and promotions by model providers, offsetting each other at the surface level. What enterprises see is a decline in headline unit prices, but an increase in actual total spending.

Unit Prices Fall, Bills Rise: The Paradox of Token Economics

Unit token prices have fallen by approximately 280 times over two years, yet enterprise AI bills have risen by an average of approximately 320%. The difference lies not in unit price, but in the structure of usage.

As enterprises move from simple chatbots to agentic workflows, a single task can consume 5 to 30 times more tokens than an ordinary conversation. Goldman Sachs estimates that agentification could push overall token demand up by 24 times.

Real cases have already emerged. Uber’s CTO confirmed that the company exhausted its entire 2026 AI coding tool budget in just four months. Behind this was the adoption of Claude Code across its approximately 5,000-person engineering team, with penetration rising from 32% to 84% and monthly API costs per employee ranging from USD 500 to USD 2,000. The FinOps Foundation’s 2026 report states that 73% of enterprises have exceeded their initial AI cost budgets.

When the primary cause of uncontrolled spending is usage structure rather than unit price, enterprises need real-time cost attribution by team, project, model, and vendor. They also need precise quota controls for high-cost models, ensuring that expensive inference is used only on tasks with clear returns, while downgradeable traffic is routed to cheaper models.

AgentsFlare records actual inference costs across four dimensions: endpoint, user, team, and model. It also supports dynamic routing based on cost budgets and SLA requirements. Once enterprises accumulate scaled usage, model billing tiers often become more flexible. The actual pricing room obtained through a unified control layer is often greater than what appears on public price lists.

95% of Pilots Deliver No Return: Enterprise AI’s ROI Problem Moves to Center Stage

Several data points are being repeatedly cited. MIT research found that 95% of generative AI pilots have produced no measurable profit-and-loss impact. BCG’s AI at Scale survey of 1,800 executives shows that only 26% of companies have achieved substantial financial value from AI investments. S&P Global reports that 42% of companies abandoned most of their AI projects in 2025.

Multiple analyses point to integration and organizational capability, rather than model quality itself, as the root issue. Enterprises that have generated returns first built three layers of foundation beneath the technology: metrics that can verify whether AI tasks are effective, infrastructure that connects tasks into automated workflows, and strategies that allow systems to continuously learn.

Part of the reason enterprise AI returns are difficult to realize is that the cost side lacks an observable and governable foundation. Pilots are either stopped because bills run out of control, or they cannot prove value because costs cannot be attributed.

For CxOs, the ROI problem is first an engineering problem: whether the organization can clearly see costs, outcomes, and invocation chains. Only after that does it become a model selection problem. An access and governance layer independent of any single model provider, offering request-level auditing, cost attribution, and quality evaluation linked to quotas, is one of the prerequisites for moving pilots into measurable production.

China’s Application Layer Splits into Two Poles: Yanyu Technology Sets a Funding Record, Moonshot AI’s Valuation Rises Sixfold in Half a Year

Yanyu Technology completed a nearly USD 300 million Series B+ financing round, reaching a post-money valuation of more than USD 2 billion. The round was jointly led by Granite Asia, Tencent, and Shunwei Capital, with follow-on participation from existing shareholders including Ant Group, Source Code Capital, and Sequoia China. The company is led by post-1990s former ByteDance employees and operates products including the AI image creation community LiblibAI, the AI video platform LibTV, and the AI design agent Xingliu.

As of May, ARR had exceeded USD 300 million, nearly tripling from the time this round was completed. Group revenue in May increased by more than 3,000% year-on-year. Revenue mainly comes from real budgets of professional content producers in short drama, film and television, advertising, and related sectors. This set a new domestic record for a single funding round in China’s AI application layer. Yanyu already has real cash flow, standing in contrast to the broader situation in which most pilots struggle to prove returns.

At the model layer, Moonshot AI, known for Kimi, reached a valuation of approximately USD 30 billion in early June, up roughly sixfold from about USD 4.3 billion at the end of 2025. In the first half of the year, it raised nearly USD 6 billion across five funding rounds, at an almost monthly cadence. After the K2.5 update, ARR increased to more than USD 200 million in April.

Capital investment in China’s AI sector has not retreated despite questions over returns, but screening criteria are tightening. Companies that can secure large financing rounds either have clear paid use cases and cash flow, or hold leading model positions with fast-growing ARR.

Application-layer vendors are rapidly stratifying, and the migration risks of being tied to a single provider are rising accordingly. During this window of rapid reshuffling across models and applications, maintaining the ability to switch among multiple vendors is more robust than betting on any current winner.

Model and Channel Layer: Gemini 3.5 Pro Nears GA, while OpenAI and Anthropic Expand Their Reach

As of mid-June, Google’s Gemini 3.5 Pro remains in limited preview for a small number of enterprise customers and is approaching general availability. Its pricing is approximately USD 15 per million input tokens and USD 60 per million output tokens, about ten times the price of Gemini 3.5 Flash. It is positioned around a 2 million-token context window and Deep Think reasoning, with the latter locked behind the USD 250-per-month Ultra tier.

OpenAI launched its Partner Network and Deployment Simulation this week, and was reported to expect losses of approximately USD 14 billion in 2026, roughly three times the level of 2025, mainly due to compute, research hiring, and infrastructure expansion.

Anthropic opened its Seoul office on June 17 and announced multiple partnerships with South Korea’s AI ecosystem. Its annualized revenue run rate reached approximately USD 47 billion in May, compared with approximately USD 4 billion 14 months earlier.

On the infrastructure side, Oracle is cutting thousands of jobs and redirecting spending toward large-scale AI data center construction, including the USD 500 billion Stargate partnership with OpenAI and SoftBank.

High-end pricing for frontier models, such as Pro-tier pricing of USD 15 / USD 60, is advancing at the same time as channel binding, including cloud credit procurement models and partner networks. The stage in which enterprises could rely on a single API is receding further.

When premium models, channel lock-in, and upstream price increases overlap, enterprises need an access layer that is not tied to any single model or cloud. Traffic should be dynamically allocated among Gemini Pro, GPT, Claude, DeepSeek, and Kimi based on task, cost, and SLA. Cost and permission policies also need to be enforced consistently across AWS, Azure, private deployments, and other environments, so that enterprises do not become passively locked in as channels and investment relationships evolve.

Overall Assessment This Week: The Bottom Layer Is Getting More Expensive, the Top Layer Has Not Yet Paid Back, and Enterprises Must Now Control Costs

At the bottom layer, memory and compute have entered a structural price increase cycle driven by supply constraints and active capacity control by manufacturers. The cost curve is unlikely to move downward in the near term. At the top layer, returns from enterprise implementation are still being seriously questioned, with most pilots unable to demonstrate profit-and-loss contribution.

These two curves are moving in opposite directions. The point of pressure falls on enterprise inference bills. The primary cause of bill escalation has shifted from unit price to usage structure. Model providers are using headline price cuts to compete for market share, while upstream suppliers are using price increases to realize profits. Enterprises sit in the middle, paying higher total bills despite falling unit prices.

This environment raises the priority of three capabilities.

First is cost observability and attribution: understanding where every dollar is spent, by team, model, and task.

Second is supplier switchability: retaining migration capability across rapidly reshuffling model and application layers, and avoiding channel lock-in.

Third is data sovereignty and cross-cloud consistency: price increases and channel consolidation often come with stronger platform lock-in, and enterprises need to ensure that their AI traffic does not pass through the infrastructure of any single vendor by default.

Together, these three capabilities form an AI control layer independent of model providers and cloud platforms. Its role has shifted from making AI easier to use to making AI affordable to use — and making its returns provable.

Enterprise Action Recommendations

Shift the billing lens from unit price to cost per outcome

Before introducing agentic workflows, establish a cost attribution baseline by team, project, model, and task. Set hard quotas for premium models, such as Gemini 3.5 Pro at USD 15 / USD 60 and Opus-tier models, and route downgradeable traffic to cheaper models. Evaluate returns based on the cost of each valid business outcome, rather than the price per million tokens. This directly addresses the problem identified by MIT and BCG: pilots that cannot prove profit-and-loss impact.

Prepare for cost pass-through from memory price increases

With memory contract prices expected to rise by more than 60% again in the second quarter and HBM sold out, cloud instance and dedicated inference pricing face upward pressure. Enterprises should lock in capacity and pricing for critical workloads while retaining the ability to route across multiple cloud and model providers, avoiding passive exposure to a single cloud’s price increase cycle. Micron’s June 24 earnings release should be closely watched for guidance on future supply and pricing.

Maintain switchability during the application-layer reshuffling window

The financing of Yanyu Technology and Moonshot AI shows that the application and model layers are rapidly stratifying. Current winners may not necessarily be long-term winners. New models and applications should be integrated through a unified control layer wherever possible, keeping migration costs manageable and avoiding irreversible deep binding to any single provider.

Treat ROI as an engineering problem first

Before expanding deployment, enterprises should complete three foundational layers: task-level measurement, workflow automation, and a continuous learning and evaluation loop. Quality evaluation should be linked to quotas and auditing. A small number of high-value scenarios should be made measurably successful before scale-up, so that enterprises do not repeat the pattern of 42% of AI projects being abandoned.