SK hynix says long-term agreements already cover more than half its production capacity for 2027 and 2028. The price is not fixed.
Generated image, not an SK hynix site
Inside this edition
Announced, committed or usable?
Six developments from 24 to 30 October 2026, with three earlier reports and studies our research desk assessed that week.
This week's lead is a chipmaker selling its future output before it exists. SK hynix, a leading maker of the memory AI systems depend on, says customers have signed up for more than half of what it can produce in 2027 and 2028.
One question runs through the issue: has something been announced, has someone committed to it, or can a customer use it today? It applies to new chips in a cloud, a security tool that fixes what it finds, a bill that is not yet law, an agent that says a task is done, and the time AI saves at work.
More than half of SK hynix's output for 2027 and 2028 is already under contract
The memory maker says long-term agreements with 14 major customers cover over half its production capacity for the next two years, at prices that move within an agreed range.
What the company said
SK hynix, the South Korean maker of memory chips used in AI data centres, published record third-quarter results on 28 October. It said it has signed long-term agreements with 14 major customers, including leading global cloud providers, and that the volume secured so far corresponds to more than half its total production capacity for 2027 and 2028. Such contracts typically run for five years, it says.
How the price works
The agreements are not fixed-price deals. SK hynix describes a mechanism that reflects market conditions within a price range agreed with each customer.
What the release cannot show
The contracts have not been published, and SK hynix says its results are preliminary and not independently verified. The capacity figure is a share of its own production, not of global memory supply.
SK hynix long-term agreements, as the company described them on 28 October
Share of its 2027 and 2028 capacity
>50%
Major customers signed
14
Company statement. Prices move within an agreed range; contracts not published.
Why it matters
Most buyers never deal with a memory maker. They buy cloud services, servers and software that depend on one. When much of future output is already allocated to the biggest customers, the useful question is what your own supplier has secured, and on what price terms.
SK hynix expects demand for memory to outrun supply until 2030. That is a forecast from an interested company, and an expert interviewed in October described how a shortage could turn.
The forecast
Citing the difficulty of leading-edge processes and the long lead time for new factories, SK hynix expects demand to exceed supply through 2030. It plans to increase investment significantly next year, while adjusting scale and timing to demand and profitability.
Another view
In a Korea Herald interview published on 15 October, Kim Hyung-joon, head of the Next Generation Intelligence Semiconductor Foundation, estimated today's memory shortage at 30 to 40%. He described how factories now being built could eventually produce more than demand, while arguing that chipmakers can now slow construction more flexibly than in past cycles. His figures are his own estimates.
COMMITTED
Volume
Over half of 2027 and 2028 capacity, by the company's account.
STILL MOVES
Price
Within a range agreed with each customer.
FORECAST
Shortage to 2030
The company's expectation, not a measured outcome.
Why it matters
Neither account settles what memory will cost in 2028. The release establishes something narrower: large buyers chose to commit volume early and accept prices that move. A smaller buyer told a price is protected should ask where that protection sits.
Oracle plans more kinds of AI chip in its cloud, on different timetables
Oracle says Google's TPU processors will reach its cloud from 2027, alongside Nvidia systems available now and further chips planned. Each part of the announcement has a different availability.
What Oracle announced
At its AI World event on 27 October, Oracle said Nvidia's latest NVL72 rack systems and processors are generally available in its cloud now. It plans to offer Google's TPUs, the chips Google designs for AI, from 2027, ramping to gigawatt scale over several years, which it says would make it the first hyperscaler outside Google to offer them. Cerebras systems are planned for 2027; SambaNova systems and OpenAI's Jalapeño chips have no date.
Its own disclaimer
Oracle says the article outlines general product direction, is not a commitment to deliver any functionality, and should not be relied on in making purchasing decisions.
What an analyst said
Data Center Knowledge quoted Brendan Burke of Futurum Group saying Oracle must prove the new chips compete with Nvidia-based offerings and show long-term commitment to them. That is judgement, not a measured comparison.
AVAILABLE NOW
Nvidia's latest NVL72 systems and processors
Generally available, Oracle says
PLANNED FROM 2027
Google TPUs and Cerebras systems
TPUs to ramp over several years
PLANNED, NO DATE
SambaNova systems and OpenAI's Jalapeño chips
Not a delivery commitment, by Oracle's own disclaimer
Why it matters
More chip types inside one cloud is a different choice from being able to move to another provider. Until regional availability, prices and results on real workloads are published, a 2027 plan is a reason to ask questions, not to change a buying decision.
An AI tool that fixes vulnerabilities, with approval first
Oracle says its AI Security Mesh uses OpenAI cyber models to find and fix vulnerabilities, but only within a scope the customer authorises and with approval before any fix.
In the same announcement, Oracle described AI Security Mesh, which it says uses OpenAI's cyber models to help find and fix software vulnerabilities. It names two controls: the tool works within customer-authorised scope, and remediation needs approval first. The announcement does not give availability, say how scope is enforced or how approvals are recorded, and no independent assessment has been published.
Why it matters
A tool that can change systems to fix them can change systems. The two controls Oracle names are the ones to examine: who defines the scope, and who approves each change.
A cyber bill asks whether a supplier could be replaced
The Cyber Security and Resilience Bill, amended in the House of Lords, would let regulators designate critical suppliers. One test is whether customers could get the service elsewhere.
Where the Bill stands
The Bill updates the UK's 2018 rules for essential and digital services. Parliament published the version amended at Lords Report stage, dated 26 October, and lists its Lords third reading for 2 November. It is not law and nothing in it is in force. Its Parliament page was updated again on 30 October.
The critical-supplier clause
Under proposed regulation 14H, a regulator could designate a supplier as critical when an incident there could disrupt an essential, digital or managed service with significant impact on the economy or daily life. The regulator must consider whether the customer could obtain the service from an alternative source, and the likely scale and duration of disruption. It must also weigh whether existing duties already manage the risk, and give the supplier reasons and a chance to respond.
DONE
Amended on Report
House of Lords text dated 26 October.
LISTED
Third reading
Lords sitting listed for 2 November.
NOT YET
Law in force
No Royal Assent, no commencement date.
Why it matters
The draft treats replaceability, not size, as part of what makes a supplier critical. The test travels: a second name on a supplier list is not an alternative unless it could take over the service, the data and the dependencies.
An AI agent's reply can pass while the record is wrong
Salesforce researchers tested AI agents on 555 customer-record tasks. Grading the reply alone overstated success, and one agent described a checkout correctly while recording it twice.
What was tested
CRM-Agent-Bench, posted on arXiv on 23 October by Salesforce researchers and assessed by our research desk on 27 October, tests agents on multi-turn tasks from retail returns to payments. Each task has a contract: the actions required, their order, the reply, the final records and records to leave alone. A task succeeds only if every check passes.
What they found
Across 16 systems, strict success ranged from 64.32% to 83.60%. Grading only the reply overstated strict success by 10.5 to 11.0 percentage points for the four systems compared this way. In one trace the reply, actions and order passed, but a checkout was recorded twice, doubling a loyalty balance. The leading system passed 434 of 555 tasks in all three rounds.
Its limits
The environments are simulated and the paper is not peer reviewed. Salesforce authorship makes it the original source, not independent validation of any product.
PASSED
The reply to the customer
Said the right thing
PASSED
The actions and their order
The expected steps were taken
FAILED
The final record
Checkout recorded twice, loyalty balance doubled
Why it matters
When an agent says a task is done, that is a claim about the record. Testing it means reading the record back, including what should not have changed.
AI saves workers time, but little of it becomes extra work done
A Hubstaff survey of US office workers finds most AI users save time each week, but few take on more work with it, and many spend time fixing colleagues' AI output.
What the survey found
Hubstaff, which sells time-tracking software, released its survey of 1,068 full-time US computer-based workers on 27 October; 839 use AI. Among AI users reporting savings, 76% save at least an hour a week. Only 26% take on more work with that time, 39% finish earlier and 15% say nothing changes. And 46% say the work expected of them has risen since they began using AI.
Where some of the time goes
Of the AI users, 45% spend an hour or more a week correcting colleagues' AI-generated work.
What it cannot show
These are unweighted self-reports from US workers, not measured output. Hubstaff's advice to measure actual work matches what it sells.
What US AI users say happens after AI saves them time
Take on more work with it
26%
Say expected workload rose
46%
Hubstaff survey of 839 AI users, October 2026. Self-reports, unweighted.
Why it matters
A saving counted per person can be spent by someone else. Before planning on AI capacity, measure the team's finished work, including the time spent checking it.
SHL, which sells workforce assessments, reports that some workers replaced by AI are later asked back by the same employer, and that most would want better terms.
SHL's survey of 2,036 US and UK adults in work, conducted in June and July and released on 29 October, calls it "AI rebounding". Firsthand experience was reported by 19% of the US group and 13% of the UK group; a wider 52% had encountered it directly, through someone they know or by hearing of it, so that is not a displacement rate. SHL attributes the pattern to employers automating roles before assessing skills, which is its own business. The full report has not been reviewed.
Why it matters
A role removed on an untested assumption can cost more to restore. Before automating one, write down what the person does that the system will not.
UK legal technology companies raised a record £227m in the first half of 2026, LawtechUK says, but fewer were funded and three took about 69% of the money.
What the figures show
Legal Futures reported LawtechUK's figures on 20 October; our research desk assessed them on 26 October. First-half investment reached £227m, 20% more than all of 2025, while the number of companies funded fell from 27 to 17. Orbital, Lawhive and Wordsmith AI raised around £156m between them.
What it cannot show
Funding shows where investors put money, not which products customers buy or whether they work.
UK lawtech investment, first half of 2026, as LawtechUK reports it
Raised in six months
£227m
Companies funded, down from 27
17
About 69% went to three companies. Reported by Legal Futures.
Why it matters
A record total can sit alongside a narrower market. For a firm choosing a supplier, a large round is a reason to ask about customers and renewals, not evidence of the answer.
Gains from AI help partly survived when the tool was removed
In a randomised experiment, AI help narrowed the gap between less and more educated adults on a business task. Some of the gain remained without AI, but much of the gap returned.
Guillermo Cruces and colleagues gave 1,174 adults a business problem with or without an AI assistant, then a follow-up without AI. Their NBER working paper, revised in October and assessed by our research desk on 28 October from its summary, reports that AI cut the gap between education groups from 0.548 to 0.139 standard deviations. Afterwards, less educated participants kept part of their gain, though a sizable gap returned. Working papers are not peer reviewed.
Why it matters
Performance with the tool and capability without it are different measures. Training that claims to build skill should test both.
Every story this week separated what was announced, what was committed and what could be used. This exercise practises that sorting with invented supplier statements.
Paper only. Use invented materials. No live assistant, accounts or real data.
01
Write six invented statements
Draft six one-line statements from a fictional supplier: a roadmap date, a signed allocation, a feature in general release, a draft rule, a survey finding and an agent's report that a job is done.
02
Sort them
Place each card under announced, committed or usable today. Usable means a customer could rely on it now, in their region, at a known price.
03
Name what would move each card
For every card not yet usable, write the one document or test that would move it: a contract clause, a published price, a read-back of the record.
04
Decide what to rely on
Finish with one line: which cards the group would plan on today, and which it would only watch.
AnnouncedCommittedUsable today
The exercise tests how a claim is read, not any real product or provider.
Sources
The evidence behind this issue
Memory, computing choice and the cyber bill text, with what each source can and cannot support.
Tier 1B · Company results release · 28 October 2026
Proposed critical-supplier provisions read, not the whole Bill. Not law; may change before Royal Assent.
Tier 1A: government, regulator, Parliament or peer-reviewed research. Tier 1B: original publication by the organisation or researchers concerned. Tier 2: established news reporting. A company announcement shows what was said, not that it works. Background stories are earlier work our research desk assessed during the week, labelled with their own dates. Sources were revisited on 31 October 2026 to prepare this edition; nothing after 30 October is reported as known that week. The Bill's Parliament page changed on 30 October; the current version is used. The cover is a generated image, not a photograph of any SK hynix site. Diagrams are original TechBridge AI artwork. Agent portraits come from the TechBridge AI organisation chart.
Sources, continued
The evidence behind this issue, continued
The Bill's progress, agent testing, work and the legal market, with what each source can and cannot support.
Tier 1A · Parliament record · Page updated 30 October 2026
Abstract only; full paper not reviewed. Not peer reviewed. An online experiment outside firms.
Until next week
Ask what you can use today.
Nine stories, one question.
SK hynix's customers committed early, at prices that still move. Oracle says its latest Nvidia systems are generally available now, while Google TPUs and Cerebras systems are planned for 2027. The cyber bill is amended, not enacted. An agent's reply is a claim until the record is read back. Time saved by AI counts only when it becomes finished work.
ONE QUESTION FOR YOUR NEXT DISCUSSION
Is this announced, committed, or ready to use today?
