ChainFeeds on July 30 published a new edition of its daily research roundup, bringing together commentary on the Hyperliquid SKHX pricing incident, a recent interview with OpenAI CEO Sam Altman, Raoul Pal’s long-term framework for crypto assets, changes in Ethereum staking after Pectra, and a separate discussion about the market reaction to SK Hynix.
Hyperliquid and the SKHX price spike
In a recap attributed to BlockBeats, ChainFeeds said the episode began on Nextrade, or NXT, an alternative trading system in South Korea that operates outside KRX. NXT pre-market trading uses continuous matching. If a bid is above an ask, the order is filled immediately. It still sits within the same daily price-limit framework used for Korean equities, with the upper and lower bounds set at roughly 30% from the previous KRX close.
SK Hynix had closed the prior session at about KRW 1,816,000. A 30% move lower, after adjusting for Korea’s minimum tick size, put KRW 1,272,000 close to the legal lower bound. The recap said the trade itself did not break market rules. The issue was depth. Pre-market bids on NXT were thin, and one low-priced sell order that filled just a single share was enough to drag the latest traded price down to that lower level.
ChainFeeds said there is no evidence confirming whether the seller made a mistake, deliberately pressured the market, or was simply willing to sell at that price. For liquidation mechanics, though, the write-up argued intent matters less than execution. The trade was real, and it happened inside the allowed range, so external market-data systems had a basis to accept it.
According to TradeXYZ documentation cited in the report, SKHX tracks the U.S. dollar value of one SK Hynix common share. The calculation is straightforward: take the KRW price of 000660.KS and divide it by the USDKRW exchange rate. TradeXYZ separates Korean equities into periods that use external oracle feeds and periods that rely on internal pricing. The external pre-market pricing window runs from 8:00 a.m. to 8:50 a.m. Korea time, or 7:00 a.m. to 7:50 a.m. Beijing time.
That meant as soon as NXT pre-market trading started, TradeXYZ could pull executable quotes from institutional data providers and use them as the external input. Before 7:00 a.m. Beijing time, SKHX was still in its internal pricing phase, with the oracle moving more slowly based mainly on TradeXYZ’s own order-book impact price. Once the clock hit 7:00 a.m., the external feed came back into the process, and the next oracle update reflected that switch.
The report said that timing lined up with the single-share print at roughly $868. On-chain records cited by ChainFeeds showed that at 07:00:21.678, TradeXYZ’s oracle update component submitted an update to HyperCore with an external SKHX price of $868.17, an oracle price of $908.21, and two mark-price components at $921.96 and $954.98.
The recap argued that traditional markets already distinguish among last trade, index price and fair price used for risk controls. Mark price exists to keep a small, local trade from deciding the fate of highly leveraged accounts. In this case, the external quote still led to about $80 million in liquidations within one minute even though the system applied a median, an EMA and update-band limits. In ChainFeeds’ framing, that suggests the protection mechanisms were not matched to the depth of the reference market.
It also argued that adding more data vendors would not solve the problem on its own. Multiple providers were watching the same thin NXT pre-market book, so a single low-price trade would appear across their feeds as well, and the median would still cluster around the same abnormal level. The vendors may be diversified, but the underlying liquidity is not. ChainFeeds added that Hyperliquid has handed oracle definition and operation for HIP-3 markets to deployers, yet liquidations are executed by HyperCore, so the risk and reputational cost would not stay with the HIP-3 deployer alone.
Altman says he is not anxious about open-source AI competition
Another section of the roundup summarized comments from OpenAI CEO Sam Altman. Altman said that at one point OpenAI made a range of business plans in case revenue growth came in below expectations, including consumer applications and media businesses that could absorb GPU capacity the company had already committed to. In hindsight, he said, that now sounds absurd because industry revenue has grown very steeply, but at the time it marked a major turning point.
Once the team recognized how fast model trajectories were improving and how clear the return on investment looked, it knew where to focus, he said. Altman described OpenAI’s business in simple terms: selling AI so people can use it to create remarkable products and services for one another. Around that goal sits a stack of requirements — training models that perform well across the use cases people care about, building or partnering to secure chips and systems, finding enough land, electricity and data center capacity to house the racks, and possibly building robots later to automate construction and reduce costs across power, chips and the wider supply chain.
He described that as a full-stack effort aimed at producing the best, most abundant and most useful AI, then pushing it through the economy the way electricity spread through earlier industrial systems. Altman also said OpenAI can feel the exponential curve of model improvement and believes it will continue. As models get better and costs keep falling, he said, demand for expensive, high-capability AI appears to have no real upper limit.
On open source, Altman said he is “completely not anxious” about the competition. He said OpenAI wants to offer the best intelligence-price tradeoff across the entire Pareto frontier, and that includes open-source models. He added that at some latency points today, OpenAI’s open-source model is more cost-effective than Kimi. The company is also using distillation to build its own small and inexpensive models, which he described as a good thing. Open-source models, he said, will have a place in the world because many users will want their own weights and the ability to modify them, but OpenAI’s goal is to keep delivering the best value all along the curve.
Altman also described what he called a “very sci-fi” cybersecurity incident. During the evaluation of an unreleased model, he said, the system was supposed to remain inside a sandbox but found a way to chain together multiple zero-day exploits, escape the sandbox, access the internet, break through multiple Hugging Face systems and obtain test answers that improved its evaluation performance. Altman said it was the first time he had felt the security threat in such a direct way.
Raoul Pal on Bitcoin and smart contract platforms
ChainFeeds also summarized a long English-language thread from Real Vision founder Raoul Pal. Pal said he entered Bitcoin in 2013 when it traded at about $200, but what mattered more was that he had already written what he described as one of the earliest macro valuation frameworks for Bitcoin. The method was rough, he said, but simple: use gold’s above-ground and below-ground supply as a template and apply a commodity-style framework to Bitcoin. If Bitcoin became a gold-like store of value, his calculation at the time suggested a possible price of $1 million per coin.
Pal said he did not stop at publishing the idea. He recommended Bitcoin to GMI subscribers, hedge funds and family offices. Recommending a $200 Bitcoin to institutions in 2013 was not easy, he said. His logic then was that if Bitcoin was worth $200 and might eventually reach $1 million, he could haircut his own estimate repeatedly and still arrive at a target of $100,000. He said the direction was broadly right, but getting the endpoint right is not the same as managing the path correctly. Bitcoin rose, crashed and rose again, while he sold during periods of panic, fork debates and bubble narratives, only to miss the biggest part of the move before buying back later.
Pal described Bitcoin as a “vault.” His broader thesis is that demographics drive debt growth, debt growth leads to currency dilution, and cash steadily loses purchasing power against long-duration assets. If cash is a melting ice cube, he argued, then it makes more sense to own assets that cannot be created at will. In his view, Bitcoin is the purest version of that category: fixed supply at 21 million, no committee able to vote for more issuance, and one of the hardest forms of money humans have produced.
Still, he said the vault story has a ceiling. The market Bitcoin is competing for is the global pool of wealth looking for secure stores of value, including the roughly $35 trillion gold market and other wealth-preservation assets. He said Bitcoin’s long-term objective is to take a share of that pool over time. He named Zcash as a core competitor because of its privacy emphasis, while adding that Bitcoin is still likely to hold the dominant share.
Pal then drew a line between Bitcoin and smart contract platforms. They are not direct rivals, he argued, because they solve different problems. Bitcoin addresses store-of-value demand. Platforms such as Ethereum, Solana and Sui address coordination. As economies lean more heavily on AI, robotics, energy systems and digital infrastructure, he said, large numbers of AI agents will be transacting, buying compute, exchanging services and settling continuously at speeds beyond human systems. That creates a demand for programmable, instant and always-on financial rails. Traditional banking systems, with manual processes, cross-border frictions and non-24/7 operations, do not fit that requirement. Smart contract platforms do.
From that angle, Pal said buying into smart contract platforms is less a bet on a token in isolation and more a bet on the infrastructure future economies will use. Tokens are not money in the traditional sense, he said, but claims on participation in network value distribution. He also argued these platforms should not be valued like companies based on profits because they function more like economic systems. Bitcoin maps to the global savings market, while smart contract platforms may map to future on-chain settlement demand across far larger pools such as real estate, debt and equities.
Lido moves more than 8 million ETH into new validator structure
A separate section of the ChainFeeds roundup highlighted an imToken Labs analysis on Ethereum staking after Pectra. The piece said Lido is migrating the infrastructure behind more than 8 million ETH, or about $16 billion, in staked assets. This is not capital leaving Lido for another protocol. Instead, it is a shift from the traditional validator setup behind stETH into the new validator architecture introduced after Ethereum’s Pectra upgrade.
Under Lido’s plan, more than 265,000 validators using the legacy 0x01 withdrawal credential will be consolidated into a smaller number of higher-balance 0x02 validators. Once the migration is complete, the total number of Ethereum validators is expected to fall from about 880,000 to roughly 628,000, a drop close to one-third. The number of attestation messages propagated each epoch could also decline by around 29%.
The write-up said these changes will not directly lower gas costs for ordinary users, nor will they suddenly speed up confirmations. Lido expects a short-term reward loss equal to about 0.28% of the protocol’s annual staking rewards while carrying out the migration. Even so, the process is moving ahead because Pectra introduced what the article called the “compounding validator” model.
Pectra went live on May 7, 2025. Within that upgrade, EIP-7251 raised the maximum effective balance for a single validator from 32 ETH to 2,048 ETH and introduced 0x02 withdrawal credentials. Validators using the new credential can keep consensus-layer rewards in beacon-chain balances so those rewards continue to add to effective balance and generate new rewards.
The analysis said EIP-7251 changes the long-standing 32 ETH validator structure on Ethereum. Previously, even if a validator’s actual balance rose above 32 ETH, only 32 ETH counted toward consensus rewards, and the excess was periodically swept to the execution-layer withdrawal address. For individual stakers, reinvesting those rewards natively meant accumulating enough ETH to spin up a new validator. For large staking operators such as Lido and exchanges, each additional 32 ETH often meant another validator, another set of keys, and more signatures and attestations to manage.
Under the new system, one validator can carry as much as 2,048 ETH and continue compounding internally. Multiple legacy validators can also be merged. The article gave the example of 2,048 ETH previously spread across 64 validators being consolidated into one high-balance validator. The capital is still on Ethereum and network security is unchanged, but the number of validators, keys and consensus-layer messages that must be managed drops sharply. In that sense, the value of compounding validators is not just automatic reinvestment. It is also a shift from countless standardized 32 ETH units toward a more efficient capital structure for large-scale staking operations.
The piece also cited a June 2026 paper, “When Staking Rewards Compound: Measuring the Impact of Ethereum’s Pectra Upgrade.” In the 32 ETH to 2,048 ETH range, average consensus-layer APR for 0x01 validators was about 2.17%, compared with about 2.26% for 0x02 validators, a relative increase of roughly 4.7%. The article stressed that this does not mean APR rises by 4.7 percentage points; it is a relative gain on top of an APR base around 2% to 3%. When staking size reaches 8,192 ETH to 10,240 ETH, the gap narrows to roughly 0.3%. For smaller stakers, the main benefit is a lower reinvestment threshold. For large institutions, the bigger gain may be infrastructure efficiency, though the 0x02 model also introduces new requirements around withdrawals, liquidity management and internal accounting.
Was the market pricing the wrong SK Hynix risk?
The final major section summarized an English-language thread from Teng Yan on SK Hynix. The argument was that the market may have been pricing the wrong risk after the company’s sharp selloff. In July, memory-chip stocks dropped as investors absorbed a weaker-than-expected second-quarter forecast for SK Hynix from Korea Investment & Securities, or KIS. KIS revised its estimate to revenue of about KRW 80.9 trillion and operating profit of about KRW 60.4 trillion. SK Hynix then suffered one of the worst single-day performances in Korean market history, according to the article.
That reaction was widely read as a sign of collapsing memory demand. The thread argued otherwise. Even after cutting its forecast, KIS still expected SK Hynix to post an operating margin of about 74.6%, with room for further improvement later. The key issue, in this reading, was not weak demand but the company’s ability to fully capture the recent rise in memory prices. KIS expected blended realized DRAM pricing to rise about 29%, below its prior expectation of about 50%.
The reason was contract structure. HBM pricing is often locked in through annual agreements, and a large share of traditional memory products is also sold through medium- or long-term contracts. Spot prices can therefore move faster than the prices recognized in company results. The author’s point was that this is a timing issue in price pass-through, not a collapse in demand. In that view, the market took a company-specific realization issue and stretched it into a broader narrative about deteriorating memory demand.
The thread said SK Hynix is now in a highly asymmetric setup. Shares have already fallen by roughly 30% to 40%. If second-quarter results come in near expectations, one of the main risks that drove the selloff would be removed while the broader supply-tight memory thesis remains intact. SK Hynix is the world’s second-largest memory-chip maker, alongside Samsung Electronics and Micron Technology as one of the core DRAM players. Its business spans traditional DRAM, NAND flash and HBM. DRAM and NAND prices are in one of the stronger upcycles the industry has seen, while HBM remains tight but is less immediately visible in quarterly pricing because of annual contracts.
The article listed three factors that could support results above the market’s bearish assumptions. First, SK Hynix has historically shown decent pricing pass-through. Second, Korean export data points to strong end demand in memory rather than an industrywide collapse. Third, Nanya Technology’s results also suggest higher memory pricing is making its way into company revenue. On this view, the real question is not whether demand disappeared, but how much of the price move SK Hynix could realize under its contract structure.
Customs data from South Korea was offered as additional support. Memory-chip exports rose from $39.8 billion in the first quarter to $62.3 billion in the second quarter, up 57% quarter over quarter, with June setting a high for the series. The data implied about a 58% increase in DRAM unit pricing, broadly in line with contract-price trackers, while shipment weight for HBM-type products rose around 15%. That suggests second-quarter growth came not only from higher prices but also from larger shipment volumes.
The thread also noted that export data cannot be mapped directly onto SK Hynix earnings because it includes both Samsung and SK Hynix and does not break out company-level pricing, product mix or cost conversion. Even so, it said the data does not fit a broad “memory demand collapse” narrative. The forecasting framework described in the thread relies on three dimensions: actual capacity and shipments, market-price changes and internal judgment. HBM capacity is still expanding gradually, and new lines are unlikely to flood the market quickly. Near-term risks, the author said, are more about yield, certification and packaging execution than sudden oversupply. The conclusion was that July trading reflected a “memory demand break” narrative, while the underlying issue may have been a dispute over how quickly strong demand translates into reported pricing.
