In early August 2026, the gas pedal for AI infrastructure was fully pressed down: on one side, Kioxia and SanDisk launched 332-layer BiCS10 3D QLC NAND, claiming to refresh the storage density ceiling; Samsung presented the zHBM concept aiming for about 10 times the density of traditional HBM5 at FMS, along with the BV-NAND prototype, which has about 58% improved density over its predecessor. On the other side, Musk announced during SpaceX's first earnings call that future AI services would exclusively run on Nvidia systems, calling the Vera Rubin architecture the "best AI computing architecture," while planning for AI computing power to exceed 2 gigawatts by the end of 2026 and approach 10 gigawatts by the end of 2027. He also proposed the concept of a space data center, with the outline of the AI computing power arms race drawn particularly aggressively within two days. Almost simultaneously, the financial side supporting this arms race began to show cracks: Oracle Cloud signed a long-term leasing commitment of about $260 billion for betting on AI data centers, which was questioned by the market regarding debt pressure and credit rating downgrade risks; the hedge fund Situational Awareness, betting on the AI theme with a high-leverage strategy, suffered significant losses, with specific details not fully disclosed, enough to become a counterexample of "AI narrative + leveraged trading." According to AiCoin data, the cryptocurrency fear and greed index rebounded from 25 to 27 between August 4 and 5, with sentiments rising from "extreme fear" to "fear." In the context of intertwining frenzies in AI infrastructure and leveraged explosions, risk appetite has not truly recovered but instead seems to be corroded by this resonance of technology and debt.

332-layer NAND and the new battlefield of zHBM storage

At the same time that sentiment rapidly contracted on a macro level, storage manufacturers were physically stacking "buildings" to new heights. Around August 4, 2026, Kioxia and SanDisk jointly released BiCS10 3D QLC NAND flash memory, stacking a single chip up to 332 layers, with interface rates reaching up to 4800 MT/s, and it was called one of the highest storage density 3D NAND products officially launched to date (according to a single source). In AI training and inference scenarios, this stacking means that more parameters and larger datasets can fit into the same volume, while high-speed interfaces determine whether these parameters can be timely "fed" to accelerators instead of being locked in cold data warehouses by bandwidth bottlenecks. For data centers being torn apart by computing power demands, this high-density, high-bandwidth NAND like BiCS10 is more like a "foundation brick" for rearranging storage hierarchies; it does not directly increase FLOPS, but fights for higher utilization for every watt of computing power.

On another front, Samsung chose to directly rewrite the storage topology around accelerators. At the FMS conference, it unveiled the zHBM concept, targeting a storage density about 10 times that of traditional HBM5, while also showcasing the V10 Bonding V-NAND (BV-NAND) prototype, which has approximately 58% higher storage density compared to the previous generation V9 NAND (according to a single source). The zHBM attempts to bring "higher capacity high-bandwidth storage" closer to the computing core through more aggressive vertical stacking, while the BV-NAND continues to compress the physical space occupied by each bit in the colder storage layer. For AI accelerator manufacturers, the overlap of these two lines points to the same path: packing more addressable parameters and intermediate activations within the same rack, same power consumption, or even the same cost framework, to support larger model scales and higher concurrent inference loads. The synchronous upgrade of storage density and bandwidth does not necessarily alleviate the current anxieties around capital expenditures for AI infrastructure but technically expands the upper limit of computing power supply, leading to future AI models' scales, forms, and cost curves being increasingly constrained by these "invisible" NAND and zHBM stacking heights.

SpaceX moves computing power into space

While storage manufacturers are still stacking NAND layers one by one on the ground, Musk has already begun planning to "launch" the computing power itself into another dimension. In SpaceX's first earnings call, he locked down the company's future AI services directly on Nvidia's tech stack—announcing that they would exclusively run on Nvidia systems and publicly supporting Nvidia's Vera Rubin architecture, calling it the "best AI computing architecture." This is not an ordinary procurement contract but a pre-laid track for computing power expansion: SpaceX's target is to scale its own AI computing power to over 2 gigawatts by the end of 2026 and approach 10 gigawatts by the end of 2027, drawing a nearly doubling power curve on the timeline.

Even more radical, this curve is not limited by the land, electricity, and regulatory costs of ground data centers. Musk proposed the concept of "space data centers" in the same earnings call, aiming to turn the already established on-orbit network infrastructure of SpaceX into a new deployment space for AI computing power and data: one end involves GPU cluster planning closely tied to Nvidia's roadmap, while the other end relies on the orbital side nodes concept based on the Starlink network, with both narratives stitched together into a "from ground to space" computing power map. For the capital market, this plan, which binds rocket launches, on-orbit networks, and AI chips together, means that the future competition in AI infrastructure is no longer just about which data center is bigger or which chip is faster, but who is bolder to bet on an ultra-long, high-risk expansion path on the balance sheet.

Oracle Cloud's AI gamble pressuring the balance sheet

If binding rockets, Starlink, and GPUs is a narrative gamble, then Oracle Cloud has completed another version of "all-in" on accounting items. According to public reports, to gain the upper hand in this round of the AI infrastructure race, Oracle Cloud signed massive data center leasing agreements with multiple operators and parks, and as of the end of fiscal year 2026, the total value of these long-term leasing commitments amounts to about $260 billion (according to a single source). Legally, these contracts are closer to rigid obligations, meaning that for many years to come, Oracle Cloud's cash flow will be locked in data centers, power, and networks at a predetermined pace, with a significant portion of the balance sheet filled with "AI data centers."

Rating agencies clearly see the other side of this gamble: when the growth path of AI business is still full of uncertainty, such a large commitment will raise overall liabilities and interest expense expectations, compress buffer space, and elevate debt and cash flow pressure as repeatedly mentioned risk points. Market discussions intensified—whether Oracle Cloud is constructing a moat for the next decade or scaffolding for a new capital expenditure bubble? Against the background of global tech stock valuations already heavily relying on "long-term AI story," this company's choice is magnified into a template: as long as Oracle Cloud's balance sheet can continue to support such expansion, other players will find it hard to tell a more "conservative" story in the capital market; once its credit risk is repriced, the entire optimistic hypothesis around the AI infrastructure cycle may also be forced to reevaluate.

High-leverage AI fund explosion cools speculation

If Oracle Cloud has taken the AI narrative to the extreme on the balance sheet, then the AI-themed hedge fund Situational Awareness has pushed leverage to its limits on the trading front. Public information shows that this fund has recently suffered significant losses under a high-leverage strategy, but specific details such as the scale of losses, position structure, and the identity of the founders remain undisclosed by authoritative channels. The rumor that its assets are being packaged to be taken over by large hedge funds is currently also at the level of market speculation, lacking verifiable evidence. However, for a market with highly exuberant sentiment, these fragmented pieces are sufficient—they have quickly been tagged as "AI asset bubble" and "risk appetite out of control," becoming a typical example: under the AI narrative, risk management is often seen as a "burden" dragging down returns until volatility truly knocks on the door.

The explosion of Situational Awareness has served as a wake-up call for many institutional risk control departments: when strategies on the AI theme stack high leverage, concentration, and correlation, any tiny disturbance from the macro or industry side can be magnified on paper into a "black swan." This repricing is not limited to a single fund or a single market—investors in traditional stock and credit markets are beginning to reassess their position exposures under the AI narrative, and this kind of risk aversion sentiment often spreads along the chain of "high-volatility assets." According to AiCoin data, the cryptocurrency fear and greed index is currently at 27; although it has slightly rebounded from the previous day's 25, it remains in the "fear" range. Amidst the backdrop of the AI-themed fund explosion and credit worries of infrastructure giants, this level resembles a note on a collective contraction of risk appetite across assets.

Chain reactions from the AI arms race to cryptocurrency panic

From the intensive emergence of breakthroughs like BiCS10, zHBM, and BV-NAND in storage technology at the beginning of August, to SpaceX pushing AI computing power planning up to several gigawatts, further adding the concept of space data centers, AI infrastructure has slipped into a high-risk, high-investment stage that requires continuous cash burn to maintain leadership; mirroring this is Oracle Cloud signing commitments of about $260 billion for data center leasing to bet on this round of expansion, only to be forced by the market to reassess its balance sheet resilience. Meanwhile, high-leverage AI funds represented by Situational Awareness have encountered significant withdrawals after layering complex leverage onto this narrative, sounding a double alarm for pricing and risk management for this arms race. According to AiCoin data, although the cryptocurrency fear and greed index has rebounded from 25 to 27, it remains in the "fear" zone, and this marginal correction is insufficient to support a new round of risk appetite elevation; instead, it seems more like a reflection of the market's difficult tug-of-war between technological optimism and cash flow anxiety. Under this macro narrative, what truly deserves attention is not just how large the next capital expenditure plans are announced by AI infrastructure providers but also whether the health of tech companies' balance sheets is continuously overdrawn against the future, as well as whether sentiment metrics for risk assets such as fear and greed can stabilize and move away from the "fear" zone, because only when the pace of capital expenditures and leverage levels return to a controllable track, without AI stories unidirectionally pulling cross-asset risk preferences, can the chain of cryptocurrency panic sparked by the AI arms race truly be interrupted.

Join our community, let's discuss and get stronger together!
AiCoin exclusive Hyperliquid benefits: https://app.hyperliquid.xyz/join/AICOIN88
AiCoin exclusive Aster benefits: https://www.asterdex.com/zh-CN/referral/9C50e2
On-chain Telegram community: https://t.me/AiCoinWhaleData
On-chain community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin on-chain Twitter: https://x.com/aicoinwhaledata

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。