The Kospi triggers two circuit breakers in a row, a historic first, and SK Hynix explains why compute is going to cost more

Two days after the Kospi’s worst session since 2008, the Korean exchange did not rebound: it fell again. On July 29, the index triggered a circuit breaker for the second consecutive session, an absolute first in the history of the Korean market, forcing the government to call an emergency meeting that very evening. The same day, a few hours apart, SK Hynix published unprecedented quarterly results (a 76% operating margin) and warned, on its own earnings call, that price negotiations for the next generation of HBM memory would factor in the surge in DRAM prices. The two stories are the same story told from both ends: the market doubts the immediate profitability of AI, at the precise moment when the component that determines its real cost (high-bandwidth memory) becomes structurally more expensive. The displayed token price keeps falling; the price of what is used to produce it goes back up.

The Kospi suspends trading for the second time running, a historic first

The July 29 session at first seemed to want to correct the previous day’s: the index opened at 6’089.11 points, up 1.09%. The respite was brief. A sell-side sidecar kicked in around 10:55, then the Korea Exchange activated a circuit breaker at 12:32:32, suspending trading for twenty minutes after the Kospi stayed for more than eight minutes more than 8% below its previous close. The index ended the session at 5’663.24 points, down 5.98%, dragging the Kosdaq down with it, also suspended. This is the first time in the history of the Korean exchange that a market-wide circuit breaker has triggered two sessions in a row. Over the two days combined, South Korean market capitalisation lost around 864’500 billion won, close to $597 billion at the day’s rate (1 USD = 1’447.8 KRW on July 29). Samsung Electronics shed a further 5.23% and SK Hynix 9.61%, after the respective falls of 14.4% and 14.7% the day before.

The political response was immediate. The South Korean government called an emergency meeting known as the “F4” at 6 p.m., bringing together Deputy Prime Minister and Finance Minister Koo Yun-cheol, Bank of Korea Governor Shin Hyun-song, Financial Services Commission Chairman Lee Eog-weon and Financial Supervisory Service Governor Lee Chan-jin. The measures announced directly target one of the amplifiers identified in the fall: leveraged ETFs on Samsung and SK Hynix, very popular with retail investors. They cap individual holdings of these products at 20% of the total portfolio, raise margin deposit requirements, reserve new purchases for professional investors and reduce the permitted leverage ratios. The authorities’ message is one of containing the market mechanics rather than disowning the industrial thesis, a distinction borne out a few hours later in SK Hynix’s figures.

SK Hynix: 76% margin, and yet the stock falls

The paradox of the day rests on a single set of results. For the second quarter of 2026, SK Hynix published revenue of 79.3 trillion won (+257% year on year), about $54.8 billion, and operating profit of 60.5 trillion won (+557% year on year), about $41.8 billion: a 76% operating margin, an extreme level for a memory chipmaker. The company began volume shipments of HBM4 during the quarter, reaching, it says, the operating speeds demanded by its customers with “industry-leading power efficiency and cost competitiveness”, and finalised long-term supply agreements with around ten strategic customers.

On this nonetheless solid base, the market kept selling the stock, a sign that the doubt is not about current demand but about the investment trajectory looking toward 2027. Asked on this point during the July 29 earnings call, management was explicit about the link between the DRAM shortage and upcoming HBM prices:

“We are currently negotiating HBM supply volumes and prices with our main customers for next year […] As standard DRAM prices have risen significantly recently, these market conditions could also partly influence HBM price negotiations.”

SK Hynix, Q2 2026 earnings call

The company specifies that it sets its prices “individually with each customer”, taking into account not only DRAM prices and the supply-demand balance, but also the “resources and opportunity costs invested in HBM”. It plans to ramp up HBM4 in the second half of 2026 and to open the Yongin site (phase 1) in early 2027.

The real signal for an AI budget: memory, not the token

It is this technical paragraph, more than the stock-market fall, that matters to anyone running a compute budget. According to TrendForce, whose analysis of June 2, 2026 remains the sector’s reference, the persistent DRAM shortage gives suppliers an unprecedented balance of power: the industry consensus now expects sharply higher HBM contract prices in 2027, with a doubling judged possible. Some estimates relayed by the specialist press suggest HBM4 going from around $2 per gigabit in the second half of 2026 to $4-5 per gigabit in 2027. Another structural signal noted in early July by TrendForce: SK Hynix has reportedly removed price caps in its long-term agreements, a position that diverges from Micron’s, meaning, concretely, less contractual protection for buyers against the next increase.

This movement feeds mechanically through the cost chain. HBM goes directly into the cost price of the AI accelerators (Nvidia, AMD) that cloud providers rent; its price rise pushes up the cost of the GPU-hour, itself the origin of the token price. The rental market already illustrates the cost gap by access channel: an H100 80 GB rents for between $1.33 and about $3 an hour on specialised clouds (Vast.ai, Lambda, RunPod), but around $7 an hour at hyperscalers (Azure at $6.98, AWS at $7.00 at the last reading), a ratio close to one to five for the same hardware. The H200, better suited to memory-hungry workloads, rents for between $3.72 and over $10 an hour depending on the provider. These gaps, which already existed, are precisely what the rise in memory costs will accentuate: suppliers who have secured volume at a fixed price will keep an advantage over those buying on the spot market in 2027.

While the stock market doubts, capacity investment is not slowing

The sharpest contrast with the stock-market turmoil came, the day before, from an infrastructure deal signed far from Seoul. On July 28, AMD and data-centre operator Core Scientific announced a partnership covering more than 500 megawatts of capacity, expandable up to 2.5 gigawatts. The agreement covers precisely 529 MW of critical IT capacity spread over five sites in Texas, Oklahoma, Alabama and Georgia: 377 MW leased for AMD and 152 MW for Core Scientific’s “Neocloud” business, under 15-year leases with three five-year renewal options, with deployment from 2027. AMD also holds a reservation right on 1’925 additional MW until December 28, 2028, and on July 27 received a warrant for 30 million Core Scientific shares at $23.47. In other words: at the very moment the Korean market revises downward the valuation of the AI memory chain, a chip supplier commits fifteen years of firm electrical capacity. The two movements do not contradict each other. One concerns the price of listed shares, the other physical contracts already signed.

Comparison table: API prices as of July 30, 2026

No change in per-token pricing has been published since the July 28 reading; the week’s movement plays out upstream, on the cost of manufacturing compute rather than its selling price. On the consumer subscription side, the market stays stable around $20 a month for the standard offer (ChatGPT Plus, Claude Pro, Google AI Pro at $19.99), and Claude Sonnet 5 keeps its launch rate of $2 / $10 until August 31, 2026, before moving to $3 / $15.

Model	Input ($/M tokens)	Output ($/M tokens)	Note
Claude Fable 5 (Anthropic)	10.00	50.00	frontier; data retention required
GPT-5.6 Sol (OpenAI)	5.00	30.00	AA index 59
Claude Opus 5 (Anthropic)	5.00	25.00	#1 on the AA index (61); $2.03/task; Fast mode ×2.5 at twice the price
GPT-5.6 Terra (OpenAI)	2.50	15.00	launched July 9
Kimi K3 (Moonshot, China, open-weight)	3.00	15.00	weights opened July 27 (1.4 TB); licence unpublished
Gemini 3.1 Pro (Google)	2.00	12.00	doubles beyond 200k context
Claude Sonnet 5	2.00	10.00	launch price until August 31, then 3/15
Gemini 3.6 Flash (Google)	1.50	7.50	launched July 21; −17% tokens/task
Grok 4.5 (xAI)	2.00	6.00	coding model; EU access still restricted
GPT-5.6 Luna (OpenAI)	1.00	6.00	launched July 9
Meta Muse Spark 1.1	1.25	4.25	Meta’s first paid API; $20 of free credits, US only
LongCat-2.0 (Meituan, open-weight)	0.75	2.95	promo 0.30/1.20; MIT licence
Gemini 3.5 Flash-Lite (Google)	0.30	2.50	launched July 21; very high throughput
Grok 4.3 / 4.20 (xAI)	1.25	2.50	1M context
Mistral Large 3	0.50	1.50	December 2025 rate (−75%); Batch API at −50%
DeepSeek V4 Pro	0.435	0.87	deepseek-reasoner alias removed July 24
DeepSeek V4 Flash	0.14	0.28	1M context; deepseek-chat alias removed July 24
Ministral 3 3B (Mistral)	0.10	0.10	cheapest paid API recorded as of July 27

Rates unchanged since the July 28, 2026 reading, no change having been published on API prices this week. The GPU rates (H100/H200) quoted in the article are aggregator readings and vary significantly by provider and date; treat them as orders of magnitude, not as a single official rate.

What this means for a Swiss organisation

Three readings, in the order in which they commit a budget.

First, the per-token price is not a leading indicator, it is a lagging one. API rates did not move this week; what moved is the price of the component that makes them possible, and it is rising on a horizon of twelve to eighteen months. An organisation that budgets its AI on today’s displayed rates is budgeting on a price that does not yet include the 2027 memory increase. This is precisely the lag documented by our July 28 article on the end of “tokenmaxxing”; in the end, volume and unit cost always catch up with each other.

Next, stock-market volatility and operational strength do not tell the same story. SK Hynix posts a 76% margin on the very day its stock falls for the second consecutive session: the market does not doubt the demand for AI memory, it doubts the pace of investment needed to meet it without excess. For a compute buyer, this distinction is operational: long-term supply contracts (AMD-Core Scientific, fifteen years; SK Hynix’s HBM agreements with around ten customers) are signed while prices fall, precisely because the two time horizons do not coincide.

Finally, dependence on a single compute supplier is becoming a price risk, not just an availability risk. When the wholesale price of memory doubles, those who have locked in volume at a fixed price keep an advantage over those buying spot the following year, and nothing guarantees that a single supplier passes that increase on in the same way as another. Diversifying your models and your suppliers is no longer just a hedge against an outage or a one-off change in pricing policy: it is a hedge against a hardware cycle whose next peak has already been announced.